Upload 20 files

.gitattributes

@@ -0,0 +1,5 @@
+checkpoints/28a7cd44f5bcd3e6cc760b65c7e0d54d.ckpt filter=lfs diff=lfs merge=lfs -text
+checkpoints/56ce1a7d241dc342982f5466c4a9d7ef.ckpt filter=lfs diff=lfs merge=lfs -text
+checkpoints/624407ef8e3a2a009f9fa51f9846fe9a.ckpt filter=lfs diff=lfs merge=lfs -text
+checkpoints/64a91b25f84141d32852e75a3aec7305.ckpt filter=lfs diff=lfs merge=lfs -text
+checkpoints/65fd1f04cb4c5847d86a9ed8ba31ac1a.ckpt filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,25 @@
+MIT License
+
+Original Sybil Model:
+Copyright (c) 2022 Peter Mikhael & Jeremy Wohlwend
+
+Hugging Face Adaptation:
+Copyright (c) 2025 Aakash Tripathi
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,260 @@
+---
+license: mit
+tags:
+- medical
+- cancer
+- ct-scan
+- risk-prediction
+- healthcare
+- pytorch
+- vision
+datasets:
+- NLST
+metrics:
+- auc
+- c-index
+language:
+- en
+library_name: transformers
+pipeline_tag: image-classification
+---
+
+# Sybil - Lung Cancer Risk Prediction
+
+## 🎯 Model Description
+
+Sybil is a validated deep learning model that predicts future lung cancer risk from a single low-dose chest CT (LDCT) scan. Published in the Journal of Clinical Oncology, this model can assess cancer risk over a 1-6 year timeframe.
+
+### Key Features
+- **Single Scan Analysis**: Requires only one LDCT scan
+- **Multi-Year Prediction**: Provides risk scores for years 1-6
+- **Validated Performance**: Tested across multiple institutions globally
+- **Ensemble Approach**: Uses 5 models for robust predictions
+
+## 🚀 Quick Start
+
+### Installation
+
+```bash
+pip install huggingface-hub torch torchvision pydicom
+```
+
+### Basic Usage
+
+```python
+from huggingface_hub import snapshot_download
+import sys
+
+# Download model
+model_path = snapshot_download(repo_id="Lab-Rasool/sybil")
+sys.path.append(model_path)
+
+# Import model
+from modeling_sybil_wrapper import SybilHFWrapper
+from configuration_sybil import SybilConfig
+
+# Initialize
+config = SybilConfig()
+model = SybilHFWrapper(config)
+
+# Prepare your DICOM files (CT scan slices)
+dicom_paths = ["scan1.dcm", "scan2.dcm", ...]  # Replace with actual paths
+
+# Get predictions
+output = model(dicom_paths=dicom_paths)
+risk_scores = output.risk_scores.numpy()
+
+# Display results
+print("Lung Cancer Risk Predictions:")
+for i, score in enumerate(risk_scores):
+    print(f"Year {i+1}: {score*100:.1f}%")
+```
+
+## 📊 Example with Demo Data
+
+```python
+import requests
+import zipfile
+from io import BytesIO
+import os
+
+# Download demo DICOM files
+def get_demo_data():
+    cache_dir = os.path.expanduser("~/.sybil_demo")
+    demo_dir = os.path.join(cache_dir, "sybil_demo_data")
+
+    if not os.path.exists(demo_dir):
+        print("Downloading demo data...")
+        url = "https://www.dropbox.com/scl/fi/covbvo6f547kak4em3cjd/sybil_example.zip?rlkey=7a13nhlc9uwga9x7pmtk1cf1c&dl=1"
+        response = requests.get(url)
+
+        os.makedirs(cache_dir, exist_ok=True)
+        with zipfile.ZipFile(BytesIO(response.content)) as zf:
+            zf.extractall(cache_dir)
+
+    # Find DICOM files
+    dicom_files = []
+    for root, dirs, files in os.walk(cache_dir):
+        for file in files:
+            if file.endswith('.dcm'):
+                dicom_files.append(os.path.join(root, file))
+
+    return sorted(dicom_files)
+
+# Run demo
+from huggingface_hub import snapshot_download
+import sys
+
+# Load model
+model_path = snapshot_download(repo_id="Lab-Rasool/sybil")
+sys.path.append(model_path)
+
+from modeling_sybil_wrapper import SybilHFWrapper
+from configuration_sybil import SybilConfig
+
+# Initialize and predict
+config = SybilConfig()
+model = SybilHFWrapper(config)
+
+dicom_files = get_demo_data()
+output = model(dicom_paths=dicom_files)
+
+# Show results
+for i, score in enumerate(output.risk_scores.numpy()):
+    print(f"Year {i+1}: {score*100:.1f}% risk")
+```
+
+Expected output for demo data:
+```
+Year 1: 2.2% risk
+Year 2: 4.5% risk
+Year 3: 7.2% risk
+Year 4: 7.9% risk
+Year 5: 9.6% risk
+Year 6: 13.6% risk
+```
+
+## 📈 Performance Metrics
+
+| Dataset | 1-Year AUC | 6-Year AUC | Sample Size |
+|---------|------------|------------|-------------|
+| NLST Test | 0.94 | 0.86 | ~15,000 |
+| MGH | 0.86 | 0.75 | ~12,000 |
+| CGMH Taiwan | 0.94 | 0.80 | ~8,000 |
+
+## 🏥 Intended Use
+
+### Primary Use Cases
+- Risk stratification in lung cancer screening programs
+- Research on lung cancer prediction models
+- Clinical decision support (with appropriate oversight)
+
+### Users
+- Healthcare providers
+- Medical researchers
+- Screening program coordinators
+
+### Out of Scope
+- ❌ Diagnosis of existing cancer
+- ❌ Use with non-LDCT imaging (X-rays, MRI)
+- ❌ Sole basis for clinical decisions
+- ❌ Use outside medical supervision
+
+## 📋 Input Requirements
+
+- **Format**: DICOM files from chest CT scan
+- **Type**: Low-dose CT (LDCT)
+- **Orientation**: Axial view
+- **Order**: Anatomically ordered (abdomen → clavicles)
+- **Number of slices**: Typically 100-300 slices
+- **Resolution**: Automatically handled by model
+
+## ⚠️ Important Considerations
+
+### Medical AI Notice
+This model should **supplement, not replace**, clinical judgment. Always consider:
+- Complete patient medical history
+- Additional risk factors (smoking, family history)
+- Current clinical guidelines
+- Need for professional medical oversight
+
+### Limitations
+- Optimized for screening population (ages 55-80)
+- Best performance with LDCT scans
+- Not validated for pediatric use
+- Performance may vary with different scanner manufacturers
+
+## 📚 Citation
+
+If you use this model, please cite the original paper:
+
+```bibtex
+@article{mikhael2023sybil,
+  title={Sybil: a validated deep learning model to predict future lung cancer risk from a single low-dose chest computed tomography},
+  author={Mikhael, Peter G and Wohlwend, Jeremy and Yala, Adam and others},
+  journal={Journal of Clinical Oncology},
+  volume={41},
+  number={12},
+  pages={2191--2200},
+  year={2023},
+  publisher={American Society of Clinical Oncology}
+}
+```
+
+## 🙏 Acknowledgments
+
+This Hugging Face implementation is based on the original work by:
+- **Original Authors**: Peter G. Mikhael & Jeremy Wohlwend
+- **Institutions**: MIT CSAIL & Massachusetts General Hospital
+- **Original Repository**: [GitHub](https://github.com/reginabarzilaygroup/Sybil)
+- **Paper**: [Journal of Clinical Oncology](https://doi.org/10.1200/JCO.22.01345)
+
+## 📄 License
+
+MIT License - See [LICENSE](LICENSE) file
+
+- Original Model © 2022 Peter Mikhael & Jeremy Wohlwend
+- HF Adaptation © 2025 Aakash Tripathi
+
+## 🔧 Troubleshooting
+
+### Common Issues
+
+1. **Import Error**: Make sure to append model path to sys.path
+   ```python
+   sys.path.append(model_path)
+   ```
+
+2. **Missing Dependencies**: Install all requirements
+   ```bash
+   pip install torch torchvision pydicom sybil huggingface-hub
+   ```
+
+3. **DICOM Loading Error**: Ensure DICOM files are valid CT scans
+   ```python
+   import pydicom
+   dcm = pydicom.dcmread("your_file.dcm")  # Test single file
+   ```
+
+4. **Memory Issues**: Model requires ~4GB GPU memory
+   ```python
+   import torch
+   device = 'cuda' if torch.cuda.is_available() else 'cpu'
+   ```
+
+## 📬 Support
+
+- **HF Model Issues**: Open issue on this repository
+- **Original Model**: [GitHub Issues](https://github.com/reginabarzilaygroup/Sybil/issues)
+- **Medical Questions**: Consult healthcare professionals
+
+## 🔍 Additional Resources
+
+- [Original GitHub Repository](https://github.com/reginabarzilaygroup/Sybil)
+- [Paper (Open Access)](https://doi.org/10.1200/JCO.22.01345)
+- [NLST Dataset Information](https://cdas.cancer.gov/nlst/)
+- [Demo Data](https://github.com/reginabarzilaygroup/Sybil/releases)
+
+---
+
+**Note**: This is a research model. Always consult qualified healthcare professionals for medical decisions.

__init__.py

@@ -0,0 +1,21 @@
+"""Hugging Face Sybil model for lung cancer risk prediction"""
+
+from .configuration_sybil import SybilConfig
+from .modeling_sybil import (
+    SybilForRiskPrediction,
+    SybilPreTrainedModel,
+    SybilOutput,
+    SybilEnsemble,
+)
+from .image_processing_sybil import SybilImageProcessor
+
+__version__ = "1.0.0"
+
+__all__ = [
+    "SybilConfig",
+    "SybilForRiskPrediction",
+    "SybilPreTrainedModel",
+    "SybilOutput",
+    "SybilEnsemble",
+    "SybilImageProcessor",
+]

config.json

@@ -0,0 +1,2018 @@
+{
+  "calibrator_data": {
+    "Year1": [
+      {
+        "coef": [
+          [
+            0.854556657547542
+          ]
+        ],
+        "intercept": [
+          -1.1545482239584481
+        ],
+        "x0": [
+          -1.1360357729796648,
+          -1.1161314190688554,
+          -1.1160988999061974,
+          -1.073612876529882,
+          -1.0735402933768126,
+          -1.0280575376367502,
+          -1.0278736221139817,
+          -0.9560547696464696,
+          -0.9534790354878826,
+          -0.8734646409329012,
+          -0.872773038089129,
+          -0.7986212739447631,
+          -0.7978832178830266,
+          -0.7657858779493016,
+          -0.7617341080705604,
+          -0.7110751930073937,
+          -0.7104674300723599,
+          -0.6202510021996833,
+          -0.6177653984897673,
+          -0.5117729076924928,
+          -0.49671799022835506,
+          -0.40163472186962823,
+          -0.37576231688531325,
+          -0.36935111064413173
+        ],
+        "x_max": -0.36935111064413173,
+        "x_min": -1.1360357729796648,
+        "y0": [
+          0.0,
+          0.0,
+          0.003745318352059925,
+          0.003745318352059925,
+          0.004878048780487805,
+          0.004878048780487805,
+          0.007194244604316547,
+          0.007194244604316547,
+          0.012987012987012988,
+          0.012987012987012988,
+          0.045454545454545456,
+          0.045454545454545456,
+          0.0625,
+          0.0625,
+          0.1,
+          0.1,
+          0.125,
+          0.125,
+          0.23529411764705882,
+          0.23529411764705882,
+          0.375,
+          0.375,
+          1.0,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            0.8610539752599486
+          ]
+        ],
+        "intercept": [
+          -1.1567984290115039
+        ],
+        "x0": [
+          -1.1378705470845625,
+          -1.106519323099567,
+          -1.1065117337381025,
+          -1.0733190326687234,
+          -1.0731443298019434,
+          -0.941562854808599,
+          -0.9414871023316992,
+          -0.8306046195608187,
+          -0.8293963521574212,
+          -0.8006963509914216,
+          -0.8005737921060307,
+          -0.6439063788575624,
+          -0.6379002236311682,
+          -0.5359935368850415,
+          -0.5292345785887509,
+          -0.4602247393296057,
+          -0.44611609712109157,
+          -0.38348501673488056,
+          -0.36534548053198035,
+          -0.36046565583271617
+        ],
+        "x_max": -0.36046565583271617,
+        "x_min": -1.1378705470845625,
+        "y0": [
+          0.0,
+          0.0,
+          0.003067484662576687,
+          0.003067484662576687,
+          0.006230529595015576,
+          0.006230529595015576,
+          0.03571428571428571,
+          0.03571428571428571,
+          0.08333333333333333,
+          0.08333333333333333,
+          0.09803921568627451,
+          0.09803921568627451,
+          0.14285714285714285,
+          0.14285714285714285,
+          0.25,
+          0.25,
+          0.5,
+          0.5,
+          1.0,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            0.931170651053989
+          ]
+        ],
+        "intercept": [
+          -1.1777126799892834
+        ],
+        "x0": [
+          -1.1551989104135245,
+          -1.1404158612667903,
+          -1.1403709080380073,
+          -1.1082794886641045,
+          -1.108175942566476,
+          -1.0433254136052175,
+          -1.042702707840477,
+          -1.0140383035344585,
+          -1.0130984171672797,
+          -0.7646651111536567,
+          -0.7600911556810273,
+          -0.7312889535464594,
+          -0.7246139107321459,
+          -0.6148745839835493,
+          -0.6138490162558539,
+          -0.3805688230821349,
+          -0.3740350608528874,
+          -0.34233082063097187,
+          -0.32275895005404653
+        ],
+        "x_max": -0.32275895005404653,
+        "x_min": -1.1551989104135245,
+        "y0": [
+          0.0,
+          0.0,
+          0.002105263157894737,
+          0.002105263157894737,
+          0.0030120481927710845,
+          0.0030120481927710845,
+          0.016666666666666666,
+          0.016666666666666666,
+          0.016853932584269662,
+          0.016853932584269662,
+          0.06666666666666667,
+          0.06666666666666667,
+          0.12903225806451613,
+          0.12903225806451613,
+          0.23333333333333334,
+          0.23333333333333334,
+          0.3333333333333333,
+          0.3333333333333333,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            0.802573263743587
+          ]
+        ],
+        "intercept": [
+          -1.133514632876246
+        ],
+        "x0": [
+          -1.1145748000846425,
+          -1.001099080307295,
+          -1.0006650304322395,
+          -0.7822120240879247,
+          -0.7809762245125411,
+          -0.7602535779621472,
+          -0.7599881299259228,
+          -0.7228992892466355,
+          -0.7204894477845987,
+          -0.6758539945396886,
+          -0.6733669918445594,
+          -0.6239612713939791,
+          -0.618395903844087,
+          -0.5797944303247502,
+          -0.5795030067463254,
+          -0.4308790865818225,
+          -0.42913485044975996,
+          -0.4127336156492978
+        ],
+        "x_max": -0.4127336156492978,
+        "x_min": -1.1145748000846425,
+        "y0": [
+          0.0,
+          0.0,
+          0.018779342723004695,
+          0.018779342723004695,
+          0.06666666666666667,
+          0.06666666666666667,
+          0.08333333333333333,
+          0.08333333333333333,
+          0.125,
+          0.125,
+          0.14285714285714285,
+          0.14285714285714285,
+          0.3333333333333333,
+          0.3333333333333333,
+          0.4,
+          0.4,
+          0.5,
+          0.5
+        ]
+      },
+      {
+        "coef": [
+          [
+            0.7571081357285535
+          ]
+        ],
+        "intercept": [
+          -1.1240990711329706
+        ],
+        "x0": [
+          -1.1080936473109704,
+          -1.0478081580528156,
+          -1.0477693656166227,
+          -0.906516611811425,
+          -0.9051812314122925,
+          -0.8537131170836589,
+          -0.8530645720823533,
+          -0.809823116729157,
+          -0.8059229792340918,
+          -0.775062431397729,
+          -0.7727583513501239,
+          -0.6678702375770744,
+          -0.6576353522247389,
+          -0.6414058198676769,
+          -0.6391149846419681,
+          -0.5392171226404363,
+          -0.5379400690975253,
+          -0.43595259388335383,
+          -0.4059231304165102,
+          -0.40304212217287894
+        ],
+        "x_max": -0.40304212217287894,
+        "x_min": -1.1080936473109704,
+        "y0": [
+          0.0,
+          0.0,
+          0.005633802816901409,
+          0.005633802816901409,
+          0.023809523809523808,
+          0.023809523809523808,
+          0.05263157894736842,
+          0.05263157894736842,
+          0.09090909090909091,
+          0.09090909090909091,
+          0.17857142857142858,
+          0.17857142857142858,
+          0.2,
+          0.2,
+          0.3333333333333333,
+          0.3333333333333333,
+          0.5,
+          0.5,
+          1.0,
+          1.0
+        ]
+      }
+    ],
+    "Year2": [
+      {
+        "coef": [
+          [
+            1.062266673654182
+          ]
+        ],
+        "intercept": [
+          -1.1402943638341543
+        ],
+        "x0": [
+          -1.1132237228060184,
+          -1.0984032927059206,
+          -1.0981069262090772,
+          -1.081592725093658,
+          -1.0814958159559502,
+          -1.033709096648227,
+          -1.0337017915615265,
+          -0.9433254512205224,
+          -0.9432834568651239,
+          -0.8978076538523547,
+          -0.8964592757256059,
+          -0.7908901199818554,
+          -0.7900304149572032,
+          -0.7597781118771237,
+          -0.7575168749204351,
+          -0.6978551950375774,
+          -0.6969377457956264,
+          -0.6754802605747152,
+          -0.6743523171985429,
+          -0.5982205872578594,
+          -0.5978797570805355,
+          -0.20437640643562294,
+          -0.17221541004131058,
+          -0.1642458849457179
+        ],
+        "x_max": -0.1642458849457179,
+        "x_min": -1.1132237228060184,
+        "y0": [
+          0.0,
+          0.0,
+          0.004166666666666667,
+          0.004166666666666667,
+          0.00646551724137931,
+          0.00646551724137931,
+          0.022950819672131147,
+          0.022950819672131147,
+          0.029850746268656716,
+          0.029850746268656716,
+          0.030927835051546393,
+          0.030927835051546393,
+          0.05555555555555555,
+          0.05555555555555555,
+          0.08,
+          0.08,
+          0.1,
+          0.1,
+          0.17391304347826086,
+          0.17391304347826086,
+          0.21875,
+          0.21875,
+          1.0,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            0.9452799797637848
+          ]
+        ],
+        "intercept": [
+          -1.1056176033444525
+        ],
+        "x0": [
+          -1.0812889671248251,
+          -1.0466438113101284,
+          -1.0466141854157391,
+          -1.0190228994163042,
+          -1.0190087854754137,
+          -0.8633880083865842,
+          -0.8629823664007918,
+          -0.8348653114530901,
+          -0.8347858113708686,
+          -0.8020762603552521,
+          -0.7996074475609661,
+          -0.7839658204696691,
+          -0.7835622350062019,
+          -0.7475163850050008,
+          -0.7461899281051441,
+          -0.5425558200873033,
+          -0.5359621587675816,
+          -0.4767176382463163,
+          -0.4681188958309078,
+          -0.4115751243820749,
+          -0.37794935041320266,
+          -0.25666073454046767,
+          -0.2367468372616991,
+          -0.2313896811187538
+        ],
+        "x_max": -0.2313896811187538,
+        "x_min": -1.0812889671248251,
+        "y0": [
+          0.0,
+          0.0,
+          0.006578947368421052,
+          0.006578947368421052,
+          0.011876484560570071,
+          0.011876484560570071,
+          0.04,
+          0.04,
+          0.043478260869565216,
+          0.043478260869565216,
+          0.07692307692307693,
+          0.07692307692307693,
+          0.1,
+          0.1,
+          0.11290322580645161,
+          0.11290322580645161,
+          0.23076923076923078,
+          0.23076923076923078,
+          0.42857142857142855,
+          0.42857142857142855,
+          0.6666666666666666,
+          0.6666666666666666,
+          1.0,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            0.977569119429414
+          ]
+        ],
+        "intercept": [
+          -1.1131720692602696
+        ],
+        "x0": [
+          -1.0845931804016375,
+          -1.065066758858955,
+          -1.0649363886111944,
+          -1.0336518336720375,
+          -1.033579472521642,
+          -1.0014803065603592,
+          -1.0011882908982757,
+          -0.6795431230600604,
+          -0.674741256056814,
+          -0.6518493782261137,
+          -0.64987028801675,
+          -0.5222888178703454,
+          -0.5212121480468886,
+          -0.3052231965882647,
+          -0.29144341938190454,
+          -0.23616471543543072,
+          -0.21561761564892545
+        ],
+        "x_max": -0.21561761564892545,
+        "x_min": -1.0845931804016375,
+        "y0": [
+          0.0,
+          0.0,
+          0.002369668246445498,
+          0.002369668246445498,
+          0.004629629629629629,
+          0.004629629629629629,
+          0.02570694087403599,
+          0.02570694087403599,
+          0.1,
+          0.1,
+          0.14285714285714285,
+          0.14285714285714285,
+          0.38461538461538464,
+          0.38461538461538464,
+          0.6666666666666666,
+          0.6666666666666666,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            0.9331653645301372
+          ]
+        ],
+        "intercept": [
+          -1.1044707886310574
+        ],
+        "x0": [
+          -1.0774566054058767,
+          -1.0420772338529536,
+          -1.0419548398643057,
+          -0.9493604640979585,
+          -0.9491299707151104,
+          -0.9365174523436794,
+          -0.9364732614670692,
+          -0.7179492653874966,
+          -0.7154435675953599,
+          -0.6731615199328296,
+          -0.67142962135454,
+          -0.6270414570925323,
+          -0.6242394940967904,
+          -0.5120045680430358,
+          -0.5055336220583483,
+          -0.33179717675997744,
+          -0.328290718304015,
+          -0.2875046806008894,
+          -0.28547662806097474,
+          -0.2664066378897004
+        ],
+        "x_max": -0.2664066378897004,
+        "x_min": -1.0774566054058767,
+        "y0": [
+          0.0,
+          0.0,
+          0.007054673721340388,
+          0.007054673721340388,
+          0.029411764705882353,
+          0.029411764705882353,
+          0.041025641025641026,
+          0.041025641025641026,
+          0.1,
+          0.1,
+          0.14285714285714285,
+          0.14285714285714285,
+          0.26666666666666666,
+          0.26666666666666666,
+          0.42105263157894735,
+          0.42105263157894735,
+          0.6666666666666666,
+          0.6666666666666666,
+          0.75,
+          0.75
+        ]
+      },
+      {
+        "coef": [
+          [
+            0.8961901274568367
+          ]
+        ],
+        "intercept": [
+          -1.0982403142894093
+        ],
+        "x0": [
+          -1.0758350750074313,
+          -1.0632542041947801,
+          -1.0631388433027222,
+          -1.0490756815117528,
+          -1.0490719723622743,
+          -1.004288045432703,
+          -1.0039359266248848,
+          -0.8597483194001383,
+          -0.857685938810237,
+          -0.7810773324119306,
+          -0.7788166141512711,
+          -0.7262313576523969,
+          -0.7216147585777637,
+          -0.6730376804506044,
+          -0.6717071674683137,
+          -0.5268755303702289,
+          -0.5241638650003476,
+          -0.40591460382490485,
+          -0.4044029534761878,
+          -0.28368021375389385,
+          -0.24813428917349833,
+          -0.24472403504569196
+        ],
+        "x_max": -0.24472403504569196,
+        "x_min": -1.0758350750074313,
+        "y0": [
+          0.0,
+          0.0,
+          0.003703703703703704,
+          0.003703703703703704,
+          0.003937007874015748,
+          0.003937007874015748,
+          0.01078167115902965,
+          0.01078167115902965,
+          0.05172413793103448,
+          0.05172413793103448,
+          0.08695652173913043,
+          0.08695652173913043,
+          0.23076923076923078,
+          0.23076923076923078,
+          0.26666666666666666,
+          0.26666666666666666,
+          0.4444444444444444,
+          0.4444444444444444,
+          0.625,
+          0.625,
+          1.0,
+          1.0
+        ]
+      }
+    ],
+    "Year3": [
+      {
+        "coef": [
+          [
+            1.072854890710045
+          ]
+        ],
+        "intercept": [
+          -1.0975483040639067
+        ],
+        "x0": [
+          -1.0668657274687154,
+          -1.055840060430324,
+          -1.0556273522769775,
+          -1.036306556853323,
+          -1.0361733269949625,
+          -1.0144346304595142,
+          -1.0142514059319403,
+          -1.0076178099514084,
+          -1.0076047887161141,
+          -0.9745636399616542,
+          -0.9744364251306473,
+          -0.893484508393487,
+          -0.8934749962572068,
+          -0.8126541951262676,
+          -0.8108757933379082,
+          -0.7132392298627734,
+          -0.7109554538692253,
+          -0.6506990920367363,
+          -0.6497724980555664,
+          -0.628972765275148,
+          -0.6269619476258803,
+          -0.5500713693644852,
+          -0.5497271419386303,
+          -0.3569641318650252,
+          -0.352800853663226,
+          -0.2250071731934148,
+          -0.21752555028003018,
+          -0.15230151923890023,
+          -0.11981995587303407,
+          -0.11177099398097867
+        ],
+        "x_max": -0.11177099398097867,
+        "x_min": -1.0668657274687154,
+        "y0": [
+          0.0,
+          0.0,
+          0.008032128514056224,
+          0.008032128514056224,
+          0.014598540145985401,
+          0.014598540145985401,
+          0.015873015873015872,
+          0.015873015873015872,
+          0.01694915254237288,
+          0.01694915254237288,
+          0.03463203463203463,
+          0.03463203463203463,
+          0.04424778761061947,
+          0.04424778761061947,
+          0.06557377049180328,
+          0.06557377049180328,
+          0.08,
+          0.08,
+          0.1111111111111111,
+          0.1111111111111111,
+          0.17391304347826086,
+          0.17391304347826086,
+          0.2222222222222222,
+          0.2222222222222222,
+          0.25,
+          0.25,
+          0.4,
+          0.4,
+          1.0,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            0.9914570976903353
+          ]
+        ],
+        "intercept": [
+          -1.0790425515518447
+        ],
+        "x0": [
+          -1.0509692533832513,
+          -1.0306651082008873,
+          -1.030658061068699,
+          -1.0208099967000621,
+          -1.020741246533108,
+          -0.985112419883033,
+          -0.9850553410670821,
+          -0.975147553361191,
+          -0.974998182229142,
+          -0.8779299862351682,
+          -0.8776659773207016,
+          -0.8460688378948997,
+          -0.845974211308601,
+          -0.741678014120675,
+          -0.7412547134258067,
+          -0.48847508646284665,
+          -0.4815593234559564,
+          -0.4194206961418063,
+          -0.4104019034227454,
+          -0.35109595731867727,
+          -0.3158275575001387,
+          -0.1886139682457576,
+          -0.16210841876048143
+        ],
+        "x_max": -0.16210841876048143,
+        "x_min": -1.0509692533832513,
+        "y0": [
+          0.0,
+          0.0,
+          0.005405405405405406,
+          0.005405405405405406,
+          0.007281553398058253,
+          0.007281553398058253,
+          0.014925373134328358,
+          0.014925373134328358,
+          0.020338983050847456,
+          0.020338983050847456,
+          0.04081632653061224,
+          0.04081632653061224,
+          0.08974358974358974,
+          0.08974358974358974,
+          0.13414634146341464,
+          0.13414634146341464,
+          0.23076923076923078,
+          0.23076923076923078,
+          0.42857142857142855,
+          0.42857142857142855,
+          0.7777777777777778,
+          0.7777777777777778,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            1.0359640190491
+          ]
+        ],
+        "intercept": [
+          -1.0883499518982132
+        ],
+        "x0": [
+          -1.0552559216978283,
+          -1.0333495853697827,
+          -1.0333285407884145,
+          -1.0011969268015661,
+          -1.0011933068093923,
+          -0.9916828772639642,
+          -0.9911346684278486,
+          -0.9698102979418284,
+          -0.9697949843715242,
+          -0.9302941777878011,
+          -0.930229033365741,
+          -0.6288182656438468,
+          -0.6237295600569457,
+          -0.5994702403937844,
+          -0.5973729296189209,
+          -0.5172434380370258,
+          -0.5152343346620796,
+          -0.4833593233187604,
+          -0.4610294208915229,
+          -0.23213841177916572,
+          -0.21753550228399432,
+          -0.1589547357990222,
+          -0.13718025903664743,
+          -0.136125351637146
+        ],
+        "x_max": -0.136125351637146,
+        "x_min": -1.0552559216978283,
+        "y0": [
+          0.0,
+          0.0,
+          0.0049261083743842365,
+          0.0049261083743842365,
+          0.012345679012345678,
+          0.012345679012345678,
+          0.014925373134328358,
+          0.014925373134328358,
+          0.04516129032258064,
+          0.04516129032258064,
+          0.05263157894736842,
+          0.05263157894736842,
+          0.1,
+          0.1,
+          0.16,
+          0.16,
+          0.2222222222222222,
+          0.2222222222222222,
+          0.3181818181818182,
+          0.3181818181818182,
+          0.6666666666666666,
+          0.6666666666666666,
+          1.0,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            0.9429092225039666
+          ]
+        ],
+        "intercept": [
+          -1.0694328972087712
+        ],
+        "x0": [
+          -1.039598461979245,
+          -1.027090895221624,
+          -1.0270806700122288,
+          -1.0023414590831823,
+          -1.0022412513285868,
+          -0.9797331874122825,
+          -0.9796138218795896,
+          -0.9353377659698549,
+          -0.9347416338032845,
+          -0.90973249280169,
+          -0.9093901256737834,
+          -0.8997258380366621,
+          -0.8996811857309828,
+          -0.6788754217433769,
+          -0.6763435601389729,
+          -0.5870183789328516,
+          -0.5841871586044692,
+          -0.47078030609401256,
+          -0.4642417922579798,
+          -0.4189702554849567,
+          -0.4188905613761421,
+          -0.288691238429426,
+          -0.2851481664915613,
+          -0.24393625211769732,
+          -0.24188702320683753,
+          -0.2226179094044748
+        ],
+        "x_max": -0.2226179094044748,
+        "x_min": -1.039598461979245,
+        "y0": [
+          0.0,
+          0.0,
+          0.0026595744680851063,
+          0.0026595744680851063,
+          0.008695652173913044,
+          0.008695652173913044,
+          0.011583011583011582,
+          0.011583011583011582,
+          0.013888888888888888,
+          0.013888888888888888,
+          0.034482758620689655,
+          0.034482758620689655,
+          0.05641025641025641,
+          0.05641025641025641,
+          0.18181818181818182,
+          0.18181818181818182,
+          0.21428571428571427,
+          0.21428571428571427,
+          0.5,
+          0.5,
+          0.5714285714285714,
+          0.5714285714285714,
+          0.6666666666666666,
+          0.6666666666666666,
+          0.75,
+          0.75
+        ]
+      },
+      {
+        "coef": [
+          [
+            0.9300079387746696
+          ]
+        ],
+        "intercept": [
+          -1.0690194713018735
+        ],
+        "x0": [
+          -1.0432696190394435,
+          -1.0296239120498654,
+          -1.029574597651545,
+          -1.0151942225358568,
+          -1.0150830797862347,
+          -0.9686807739460517,
+          -0.9686333650499847,
+          -0.9202678722062687,
+          -0.9200592370322587,
+          -0.8215279605321205,
+          -0.8193877558337873,
+          -0.8032501334690247,
+          -0.8001076484435512,
+          -0.7398883171030026,
+          -0.7375422904448954,
+          -0.682972724782267,
+          -0.6781819179456279,
+          -0.6402737868235991,
+          -0.6374435270037346,
+          -0.6026054025305556,
+          -0.6024444257001804,
+          -0.4871964897395866,
+          -0.47609418826063177,
+          -0.350568790778378,
+          -0.3490000981741114,
+          -0.22372187551508194,
+          -0.18683462242218773,
+          -0.18329568206240143
+        ],
+        "x_max": -0.18329568206240143,
+        "x_min": -1.0432696190394435,
+        "y0": [
+          0.0,
+          0.0,
+          0.0040650406504065045,
+          0.0040650406504065045,
+          0.009708737864077669,
+          0.009708737864077669,
+          0.015151515151515152,
+          0.015151515151515152,
+          0.017142857142857144,
+          0.017142857142857144,
+          0.05555555555555555,
+          0.05555555555555555,
+          0.09523809523809523,
+          0.09523809523809523,
+          0.13043478260869565,
+          0.13043478260869565,
+          0.2727272727272727,
+          0.2727272727272727,
+          0.3,
+          0.3,
+          0.30434782608695654,
+          0.30434782608695654,
+          0.5,
+          0.5,
+          0.625,
+          0.625,
+          1.0,
+          1.0
+        ]
+      }
+    ],
+    "Year4": [
+      {
+        "coef": [
+          [
+            1.0713611402736163
+          ]
+        ],
+        "intercept": [
+          -1.0720420740830974
+        ],
+        "x0": [
+          -1.0414022172560469,
+          -1.0303919014036294,
+          -1.030179489406734,
+          -1.0108855945889734,
+          -1.0107525502283543,
+          -0.989044120775359,
+          -0.9888611513537834,
+          -0.9822367914190364,
+          -0.9822237883133847,
+          -0.959024555591222,
+          -0.9588443480324798,
+          -0.9492286431991185,
+          -0.9491016054910486,
+          -0.868262399207094,
+          -0.8682529003146905,
+          -0.8274793436967418,
+          -0.826701839397852,
+          -0.6882680786258673,
+          -0.6859874823646543,
+          -0.6041189393814397,
+          -0.6021109214208971,
+          -0.5253273989559846,
+          -0.5249836508026502,
+          -0.20071579408044027,
+          -0.1932445879322574,
+          -0.12811136919545385,
+          -0.09567503035144409,
+          -0.08763727513835629
+        ],
+        "x_max": -0.08763727513835629,
+        "x_min": -1.0414022172560469,
+        "y0": [
+          0.0,
+          0.0,
+          0.008130081300813007,
+          0.008130081300813007,
+          0.01486988847583643,
+          0.01486988847583643,
+          0.015873015873015872,
+          0.015873015873015872,
+          0.018292682926829267,
+          0.018292682926829267,
+          0.028169014084507043,
+          0.028169014084507043,
+          0.04329004329004329,
+          0.04329004329004329,
+          0.06896551724137931,
+          0.06896551724137931,
+          0.07964601769911504,
+          0.07964601769911504,
+          0.11764705882352941,
+          0.11764705882352941,
+          0.21739130434782608,
+          0.21739130434782608,
+          0.25,
+          0.25,
+          0.4,
+          0.4,
+          1.0,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            1.0236741286444109
+          ]
+        ],
+        "intercept": [
+          -1.0625940402455896
+        ],
+        "x0": [
+          -1.0336085106570108,
+          -1.0144313554536946,
+          -1.014352061695433,
+          -1.0024692401049637,
+          -1.0023982559268134,
+          -0.9959623202441721,
+          -0.9959512916504678,
+          -0.9667386748515856,
+          -0.9665320450724462,
+          -0.8652485370410349,
+          -0.8652359983079437,
+          -0.8220499321037427,
+          -0.821952230661591,
+          -0.7693876649771454,
+          -0.7693015717781595,
+          -0.7142669671983582,
+          -0.7138299115042861,
+          -0.7017531895419336,
+          -0.700041515189594,
+          -0.45283630424364407,
+          -0.4456958160808363,
+          -0.38153801710077406,
+          -0.3722261620546843,
+          -0.3109930912108302,
+          -0.2745786578107059,
+          -0.1432313101190359,
+          -0.1216659095169288,
+          -0.1158644726262289
+        ],
+        "x_max": -0.1158644726262289,
+        "x_min": -1.0336085106570108,
+        "y0": [
+          0.0,
+          0.0,
+          0.00975609756097561,
+          0.00975609756097561,
+          0.010752688172043012,
+          0.010752688172043012,
+          0.019543973941368076,
+          0.019543973941368076,
+          0.02586206896551724,
+          0.02586206896551724,
+          0.046875,
+          0.046875,
+          0.06976744186046512,
+          0.06976744186046512,
+          0.09090909090909091,
+          0.09090909090909091,
+          0.14285714285714285,
+          0.14285714285714285,
+          0.14864864864864866,
+          0.14864864864864866,
+          0.23076923076923078,
+          0.23076923076923078,
+          0.42857142857142855,
+          0.42857142857142855,
+          0.7777777777777778,
+          0.7777777777777778,
+          1.0,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            1.0396994880608152
+          ]
+        ],
+        "intercept": [
+          -1.0658926889834355
+        ],
+        "x0": [
+          -1.0326793286499445,
+          -1.0106940026668443,
+          -1.0106728822031288,
+          -0.9784254083542221,
+          -0.9784217753091158,
+          -0.9469256058925717,
+          -0.9469102371047443,
+          -0.9370942140091818,
+          -0.9369620532787766,
+          -0.915314170567322,
+          -0.9148419676598699,
+          -0.8716363389381502,
+          -0.870683567049227,
+          -0.6047040278319293,
+          -0.5995969734397801,
+          -0.5752501797570768,
+          -0.5731453065193639,
+          -0.49272688479392623,
+          -0.49071053701351774,
+          -0.4587205910634634,
+          -0.43631017169063924,
+          -0.20659382959771744,
+          -0.19193826507312706,
+          -0.1331462686285585,
+          -0.11129327766907215
+        ],
+        "x_max": -0.11129327766907215,
+        "x_min": -1.0326793286499445,
+        "y0": [
+          0.0,
+          0.0,
+          0.009950248756218905,
+          0.009950248756218905,
+          0.014150943396226415,
+          0.014150943396226415,
+          0.045454545454545456,
+          0.045454545454545456,
+          0.05,
+          0.05,
+          0.0673076923076923,
+          0.0673076923076923,
+          0.07142857142857142,
+          0.07142857142857142,
+          0.1,
+          0.1,
+          0.2,
+          0.2,
+          0.2222222222222222,
+          0.2222222222222222,
+          0.4090909090909091,
+          0.4090909090909091,
+          0.6666666666666666,
+          0.6666666666666666,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            0.9866801376021137
+          ]
+        ],
+        "intercept": [
+          -1.055717348217441
+        ],
+        "x0": [
+          -1.0244979646808392,
+          -1.0114097824943242,
+          -1.0113990826191164,
+          -0.9855114494715504,
+          -0.9854065899593796,
+          -0.9761218403287525,
+          -0.9760201492440737,
+          -0.9512964316870484,
+          -0.9511816258120083,
+          -0.8886034699831773,
+          -0.8882452097850468,
+          -0.8781322949592707,
+          -0.8780855698429388,
+          -0.7753388033507341,
+          -0.7726365096184081,
+          -0.6674841788771404,
+          -0.6664703408904253,
+          -0.6470297519604127,
+          -0.644380358469925,
+          -0.550908601082995,
+          -0.5479459522971747,
+          -0.429274625122231,
+          -0.422432586061124,
+          -0.2387327764997047,
+          -0.23502523114565355,
+          -0.19190021298402493,
+          -0.18975585654388327,
+          -0.16959224863314815
+        ],
+        "x_max": -0.16959224863314815,
+        "x_min": -1.0244979646808392,
+        "y0": [
+          0.0,
+          0.0,
+          0.008152173913043478,
+          0.008152173913043478,
+          0.010416666666666666,
+          0.010416666666666666,
+          0.015151515151515152,
+          0.015151515151515152,
+          0.019230769230769232,
+          0.019230769230769232,
+          0.034482758620689655,
+          0.034482758620689655,
+          0.056910569105691054,
+          0.056910569105691054,
+          0.07936507936507936,
+          0.07936507936507936,
+          0.2,
+          0.2,
+          0.20588235294117646,
+          0.20588235294117646,
+          0.26666666666666666,
+          0.26666666666666666,
+          0.5,
+          0.5,
+          0.6666666666666666,
+          0.6666666666666666,
+          0.75,
+          0.75
+        ]
+      },
+      {
+        "coef": [
+          [
+            0.9720870020574502
+          ]
+        ],
+        "intercept": [
+          -1.0561492900363496
+        ],
+        "x0": [
+          -1.0292343620603719,
+          -1.0149712425262198,
+          -1.0149196968526584,
+          -0.9998886684583321,
+          -0.9997724969529331,
+          -0.9634315079071001,
+          -0.9634086791923234,
+          -0.951270676170449,
+          -0.9512211222153406,
+          -0.8147576198447524,
+          -0.8126328401139585,
+          -0.7783549744382297,
+          -0.7750703047918094,
+          -0.7121262945010313,
+          -0.7096741197122537,
+          -0.6526355039360778,
+          -0.6476279326443739,
+          -0.4480011933435484,
+          -0.43639655803813926,
+          -0.3051916486678352,
+          -0.3035519791260457,
+          -0.17260542823084102,
+          -0.13404917749555145,
+          -0.13035011452912415
+        ],
+        "x_max": -0.13035011452912415,
+        "x_min": -1.0292343620603719,
+        "y0": [
+          0.0,
+          0.0,
+          0.00819672131147541,
+          0.00819672131147541,
+          0.011467889908256881,
+          0.011467889908256881,
+          0.012987012987012988,
+          0.012987012987012988,
+          0.036211699164345405,
+          0.036211699164345405,
+          0.07142857142857142,
+          0.07142857142857142,
+          0.075,
+          0.075,
+          0.13043478260869565,
+          0.13043478260869565,
+          0.3170731707317073,
+          0.3170731707317073,
+          0.5,
+          0.5,
+          0.7142857142857143,
+          0.7142857142857143,
+          1.0,
+          1.0
+        ]
+      }
+    ],
+    "Year5": [
+      {
+        "coef": [
+          [
+            1.1130009773475025
+          ]
+        ],
+        "intercept": [
+          -1.05794336672334
+        ],
+        "x0": [
+          -1.0261126522122785,
+          -1.0146744061867454,
+          -1.0144537385227264,
+          -0.9969480439120526,
+          -0.9968250495003456,
+          -0.9737491227404367,
+          -0.9736590578600733,
+          -0.9412351121614915,
+          -0.940990193071027,
+          -0.9329694846791076,
+          -0.9328188264757706,
+          -0.8462435476743763,
+          -0.8462336630102207,
+          -0.8038753893527305,
+          -0.8030676331731484,
+          -0.6592534983234417,
+          -0.6568842637374115,
+          -0.6446790254063086,
+          -0.6381987988020102,
+          -0.5718338270620633,
+          -0.5697477317150653,
+          -0.4899798519505131,
+          -0.48962280992041207,
+          -0.15275186333929358,
+          -0.14499027909631035,
+          -0.07732563986998098,
+          -0.04362862074369689,
+          -0.03527840143325056
+        ],
+        "x_max": -0.03527840143325056,
+        "x_min": -1.0261126522122785,
+        "y0": [
+          0.0,
+          0.0,
+          0.009950248756218905,
+          0.009950248756218905,
+          0.018867924528301886,
+          0.018867924528301886,
+          0.020242914979757085,
+          0.020242914979757085,
+          0.05084745762711865,
+          0.05084745762711865,
+          0.05150214592274678,
+          0.05150214592274678,
+          0.07017543859649122,
+          0.07017543859649122,
+          0.09433962264150944,
+          0.09433962264150944,
+          0.14285714285714285,
+          0.14285714285714285,
+          0.14814814814814814,
+          0.14814814814814814,
+          0.21739130434782608,
+          0.21739130434782608,
+          0.28,
+          0.28,
+          0.4,
+          0.4,
+          1.0,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            1.0638269123140547
+          ]
+        ],
+        "intercept": [
+          -1.0500459135423696
+        ],
+        "x0": [
+          -1.0199234501712975,
+          -0.9999940866794924,
+          -0.9999116787251026,
+          -0.9875627671022001,
+          -0.9874889986274953,
+          -0.9504306995378152,
+          -0.9502159648739382,
+          -0.9230139844783734,
+          -0.922969931058108,
+          -0.8449597096117958,
+          -0.8449466632048586,
+          -0.8000666587391363,
+          -0.7999651250375173,
+          -0.7453387570904597,
+          -0.7452492869558871,
+          -0.6880559631457261,
+          -0.6876018277071652,
+          -0.4163708637296526,
+          -0.4089503591453124,
+          -0.3422760227720679,
+          -0.3325989812075447,
+          -0.26896409304876334,
+          -0.2311212697577245,
+          -0.14109465757978035,
+          -0.1050875889776246,
+          -0.066181523151848
+        ],
+        "x_max": -0.066181523151848,
+        "x_min": -1.0199234501712975,
+        "y0": [
+          0.0,
+          0.0,
+          0.010101010101010102,
+          0.010101010101010102,
+          0.02570694087403599,
+          0.02570694087403599,
+          0.02857142857142857,
+          0.02857142857142857,
+          0.031088082901554404,
+          0.031088082901554404,
+          0.08196721311475409,
+          0.08196721311475409,
+          0.0975609756097561,
+          0.0975609756097561,
+          0.11764705882352941,
+          0.11764705882352941,
+          0.14285714285714285,
+          0.14285714285714285,
+          0.3076923076923077,
+          0.3076923076923077,
+          0.42857142857142855,
+          0.42857142857142855,
+          0.8333333333333334,
+          0.8333333333333334,
+          1.0,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            1.096758407454776
+          ]
+        ],
+        "intercept": [
+          -1.055167618027271
+        ],
+        "x0": [
+          -1.020131505808367,
+          -1.0059211960996113,
+          -1.0059050165555292,
+          -0.996939616577453,
+          -0.9969173370183926,
+          -0.9835280206852376,
+          -0.9833181850728733,
+          -0.929671597216839,
+          -0.9296553849868093,
+          -0.9193006483755612,
+          -0.919161242808875,
+          -0.8963253198907385,
+          -0.8958272187346468,
+          -0.8878365014579762,
+          -0.8877675667944027,
+          -0.8502504399770857,
+          -0.8492453634301178,
+          -0.5686688293396501,
+          -0.5632815641355099,
+          -0.5375985154051841,
+          -0.5353781589819728,
+          -0.45054635791402964,
+          -0.4484194179291747,
+          -0.4146738572381863,
+          -0.3910334841196985,
+          -0.2983852309049294,
+          -0.29808164382324565,
+          -0.14871026244525198,
+          -0.13325039755840862,
+          -0.07123188438986094,
+          -0.048179596841319006
+        ],
+        "x_max": -0.048179596841319006,
+        "x_min": -1.020131505808367,
+        "y0": [
+          0.0,
+          0.0,
+          0.007518796992481203,
+          0.007518796992481203,
+          0.017543859649122806,
+          0.017543859649122806,
+          0.021479713603818614,
+          0.021479713603818614,
+          0.04878048780487805,
+          0.04878048780487805,
+          0.06329113924050633,
+          0.06329113924050633,
+          0.06896551724137931,
+          0.06896551724137931,
+          0.07142857142857142,
+          0.07142857142857142,
+          0.07453416149068323,
+          0.07453416149068323,
+          0.1111111111111111,
+          0.1111111111111111,
+          0.2,
+          0.2,
+          0.25,
+          0.25,
+          0.42857142857142855,
+          0.42857142857142855,
+          0.4375,
+          0.4375,
+          0.6666666666666666,
+          0.6666666666666666,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            1.0192893240958014
+          ]
+        ],
+        "intercept": [
+          -1.0434135518718168
+        ],
+        "x0": [
+          -1.0111623826452667,
+          -0.9959888476152152,
+          -0.9959373241054497,
+          -0.9611874678027185,
+          -0.9610824158887104,
+          -0.9355415867075108,
+          -0.9354229865671895,
+          -0.8979901713623936,
+          -0.8978116598109821,
+          -0.8846290685337032,
+          -0.8843235043891193,
+          -0.8755302346654252,
+          -0.8751063969353022,
+          -0.8599594191627716,
+          -0.8599111498093823,
+          -0.7541104380913812,
+          -0.7509770614187808,
+          -0.6423494788656237,
+          -0.6413021645189869,
+          -0.6212190757900966,
+          -0.6184820606801298,
+          -0.52192117855335,
+          -0.5188606160028857,
+          -0.39626727267313655,
+          -0.3891991083266022,
+          -0.34026030244508254,
+          -0.3401742134913346,
+          -0.19942813028762418,
+          -0.19559799202994355,
+          -0.1510479001582279,
+          -0.14883255251569716,
+          -0.1280025494626167
+        ],
+        "x_max": -0.1280025494626167,
+        "x_min": -1.0111623826452667,
+        "y0": [
+          0.0,
+          0.0,
+          0.0136986301369863,
+          0.0136986301369863,
+          0.021739130434782608,
+          0.021739130434782608,
+          0.02247191011235955,
+          0.02247191011235955,
+          0.02564102564102564,
+          0.02564102564102564,
+          0.05,
+          0.05,
+          0.05405405405405406,
+          0.05405405405405406,
+          0.059322033898305086,
+          0.059322033898305086,
+          0.109375,
+          0.109375,
+          0.2,
+          0.2,
+          0.2727272727272727,
+          0.2727272727272727,
+          0.2857142857142857,
+          0.2857142857142857,
+          0.5,
+          0.5,
+          0.6,
+          0.6,
+          0.6666666666666666,
+          0.6666666666666666,
+          0.75,
+          0.75
+        ]
+      },
+      {
+        "coef": [
+          [
+            1.0246281291281125
+          ]
+        ],
+        "intercept": [
+          -1.046126360981268
+        ],
+        "x0": [
+          -1.017756686004363,
+          -1.0076691078074915,
+          -1.007111022611281,
+          -0.9865689458798288,
+          -0.9864557783600643,
+          -0.9355790769588279,
+          -0.9355268446213867,
+          -0.8592524325149089,
+          -0.8591646101223447,
+          -0.8017786006234253,
+          -0.8017372850124589,
+          -0.7916875145693405,
+          -0.7894478908774645,
+          -0.6835089846082325,
+          -0.6809242702145596,
+          -0.6208027275469761,
+          -0.6155244368578598,
+          -0.5322587317287224,
+          -0.5320813769109154,
+          -0.40510803046102895,
+          -0.392876105638554,
+          -0.25457965566226204,
+          -0.2528513622775791,
+          -0.11482711286015956,
+          -0.07418678158401704,
+          -0.07028790707685573
+        ],
+        "x_max": -0.07028790707685573,
+        "x_min": -1.017756686004363,
+        "y0": [
+          0.0,
+          0.0,
+          0.010638297872340425,
+          0.010638297872340425,
+          0.016393442622950817,
+          0.016393442622950817,
+          0.04918032786885246,
+          0.04918032786885246,
+          0.05555555555555555,
+          0.05555555555555555,
+          0.07142857142857142,
+          0.07142857142857142,
+          0.08333333333333333,
+          0.08333333333333333,
+          0.13043478260869565,
+          0.13043478260869565,
+          0.3157894736842105,
+          0.3157894736842105,
+          0.3333333333333333,
+          0.3333333333333333,
+          0.5,
+          0.5,
+          0.7142857142857143,
+          0.7142857142857143,
+          1.0,
+          1.0
+        ]
+      }
+    ],
+    "Year6": [
+      {
+        "coef": [
+          [
+            1.340606917145108
+          ]
+        ],
+        "intercept": [
+          -1.0475421908007814
+        ],
+        "x0": [
+          -1.0092021741074475,
+          -0.995633400824669,
+          -0.9951590414806812,
+          -0.9820283475396332,
+          -0.9817901515581261,
+          -0.906122027838656,
+          -0.9058965319804526,
+          -0.8970114597943656,
+          -0.8968299923620552,
+          -0.7925502837390519,
+          -0.792538377685589,
+          -0.7439244631515856,
+          -0.7405450017919006,
+          -0.5673212330021985,
+          -0.5644674958186597,
+          -0.46202445575510687,
+          -0.4595117590828409,
+          -0.35183289217277947,
+          -0.35087936911255946,
+          0.04275873029063826,
+          0.05210753926378531,
+          0.1336094294140786,
+          0.17419740538830197,
+          0.18425522384048465
+        ],
+        "x_max": 0.18425522384048465,
+        "x_min": -1.0092021741074475,
+        "y0": [
+          0.0,
+          0.0,
+          0.022988505747126436,
+          0.022988505747126436,
+          0.03211009174311927,
+          0.03211009174311927,
+          0.046511627906976744,
+          0.046511627906976744,
+          0.06936416184971098,
+          0.06936416184971098,
+          0.10810810810810811,
+          0.10810810810810811,
+          0.13414634146341464,
+          0.13414634146341464,
+          0.18518518518518517,
+          0.18518518518518517,
+          0.3333333333333333,
+          0.3333333333333333,
+          0.35294117647058826,
+          0.35294117647058826,
+          0.5,
+          0.5,
+          1.0,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            1.2907222572139896
+          ]
+        ],
+        "intercept": [
+          -1.039821376539258
+        ],
+        "x0": [
+          -1.0032743278933614,
+          -0.9791775889843037,
+          -0.9789944017955169,
+          -0.9643426097971722,
+          -0.963922185149528,
+          -0.9613656665174612,
+          -0.961356819217672,
+          -0.9205582848036054,
+          -0.9202258763639181,
+          -0.7909939357247382,
+          -0.7909781067514198,
+          -0.7780041955414604,
+          -0.7776605356444324,
+          -0.7365259961709816,
+          -0.7364028071380476,
+          -0.7206493820664985,
+          -0.7201010418121743,
+          -0.6034922842627248,
+          -0.6000744184880863,
+          -0.27099471764664695,
+          -0.26199155151510733,
+          -0.1810967659576298,
+          -0.16935578367307202,
+          -0.09214870653380258,
+          -0.04623468261490005,
+          0.06299300351957249,
+          0.10667973894981264,
+          0.15388377611101745
+        ],
+        "x_max": 0.15388377611101745,
+        "x_min": -1.0032743278933614,
+        "y0": [
+          0.0,
+          0.0,
+          0.014388489208633094,
+          0.014388489208633094,
+          0.04,
+          0.04,
+          0.0421455938697318,
+          0.0421455938697318,
+          0.05138339920948617,
+          0.05138339920948617,
+          0.07692307692307693,
+          0.07692307692307693,
+          0.08333333333333333,
+          0.08333333333333333,
+          0.14285714285714285,
+          0.14285714285714285,
+          0.15217391304347827,
+          0.15217391304347827,
+          0.20689655172413793,
+          0.20689655172413793,
+          0.36363636363636365,
+          0.36363636363636365,
+          0.6,
+          0.6,
+          0.8333333333333334,
+          0.8333333333333334,
+          1.0,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            1.2981443067116996
+          ]
+        ],
+        "intercept": [
+          -1.0424440037259006
+        ],
+        "x0": [
+          -1.0009745886777577,
+          -0.9841549932959078,
+          -0.9841358428769162,
+          -0.9735242233444598,
+          -0.9734978528306313,
+          -0.9576500009696335,
+          -0.9574016355103996,
+          -0.893904504212875,
+          -0.8938853151061683,
+          -0.8816292500621034,
+          -0.8814642469570553,
+          -0.8544352101854283,
+          -0.8538456480944956,
+          -0.8443876817194529,
+          -0.8443060893282346,
+          -0.79990009777071,
+          -0.7987104698741856,
+          -0.7891257045127901,
+          -0.7888712360664938,
+          -0.42983927753286244,
+          -0.4272112210437845,
+          -0.3268026726926374,
+          -0.32428518569341347,
+          -0.284343292209814,
+          -0.2563620927371575,
+          -0.14670184156420663,
+          -0.1463425100660386,
+          0.030456324575673266,
+          0.048754917455359426,
+          0.11314767903017375,
+          0.1494463503795691
+        ],
+        "x_max": 0.1494463503795691,
+        "x_min": -1.0009745886777577,
+        "y0": [
+          0.0,
+          0.0,
+          0.00980392156862745,
+          0.00980392156862745,
+          0.022556390977443608,
+          0.022556390977443608,
+          0.03289473684210526,
+          0.03289473684210526,
+          0.0625,
+          0.0625,
+          0.08771929824561403,
+          0.08771929824561403,
+          0.1,
+          0.1,
+          0.10526315789473684,
+          0.10526315789473684,
+          0.1111111111111111,
+          0.1111111111111111,
+          0.14545454545454545,
+          0.14545454545454545,
+          0.2631578947368421,
+          0.2631578947368421,
+          0.2857142857142857,
+          0.2857142857142857,
+          0.42857142857142855,
+          0.42857142857142855,
+          0.6363636363636364,
+          0.6363636363636364,
+          0.8,
+          0.8,
+          1.0
+        ]
+      },
+      {
+        "coef": [
+          [
+            1.267499953639535
+          ]
+        ],
+        "intercept": [
+          -1.035861815812616
+        ],
+        "x0": [
+          -0.9957570540603671,
+          -0.9821540666109716,
+          -0.982102731143961,
+          -0.9638658914809329,
+          -0.9638377872959027,
+          -0.9336125767773732,
+          -0.9334819433124802,
+          -0.8550258905297626,
+          -0.8548039090199113,
+          -0.8384111798528299,
+          -0.8380312067382899,
+          -0.8279540810062146,
+          -0.826569609923502,
+          -0.8077341397870003,
+          -0.8076741161982733,
+          -0.6761095113040186,
+          -0.6722131153609145,
+          -0.5384220470663117,
+          -0.5358309092005328,
+          -0.5108573188459119,
+          -0.5074538038252172,
+          -0.33549109502565666,
+          -0.3195973473534073,
+          -0.2311267327585821,
+          -0.22233737553707267,
+          -0.16148131365196827,
+          -0.16137426088264661,
+          0.0738068478711209,
+          0.07656166239514639,
+          0.10246405073568132
+        ],
+        "x_max": 0.10246405073568132,
+        "x_min": -0.9957570540603671,
+        "y0": [
+          0.0,
+          0.0,
+          0.024691358024691357,
+          0.024691358024691357,
+          0.0297029702970297,
+          0.0297029702970297,
+          0.03543307086614173,
+          0.03543307086614173,
+          0.04,
+          0.04,
+          0.06666666666666667,
+          0.06666666666666667,
+          0.08,
+          0.08,
+          0.08791208791208792,
+          0.08791208791208792,
+          0.13725490196078433,
+          0.13725490196078433,
+          0.2857142857142857,
+          0.2857142857142857,
+          0.2972972972972973,
+          0.2972972972972973,
+          0.3333333333333333,
+          0.3333333333333333,
+          0.5714285714285714,
+          0.5714285714285714,
+          0.7272727272727273,
+          0.7272727272727273,
+          0.75,
+          0.75
+        ]
+      },
+      {
+        "coef": [
+          [
+            1.262442132690388
+          ]
+        ],
+        "intercept": [
+          -1.0375704296176012
+        ],
+        "x0": [
+          -1.0026162135821177,
+          -0.9901873299993725,
+          -0.9894997144193408,
+          -0.9646615162846798,
+          -0.9640504320285327,
+          -0.9013653567360134,
+          -0.9013010013844471,
+          -0.8107793489199652,
+          -0.8072152551147188,
+          -0.7464076686646566,
+          -0.7457327181651399,
+          -0.5907903335367819,
+          -0.5876057124447174,
+          -0.5135300879442599,
+          -0.5070267172431278,
+          -0.40443521953298234,
+          -0.40421670104020413,
+          -0.2537263109006871,
+          -0.232702183186429,
+          -0.06230742211331197,
+          -0.0601779955199484,
+          0.10988138782509971,
+          0.15995425331743718,
+          0.16475804828264184
+        ],
+        "x_max": 0.16475804828264184,
+        "x_min": -1.0026162135821177,
+        "y0": [
+          0.0,
+          0.0,
+          0.014563106796116505,
+          0.014563106796116505,
+          0.027932960893854747,
+          0.027932960893854747,
+          0.0755813953488372,
+          0.0755813953488372,
+          0.1016949152542373,
+          0.1016949152542373,
+          0.109375,
+          0.109375,
+          0.15,
+          0.15,
+          0.4,
+          0.4,
+          0.4375,
+          0.4375,
+          0.5,
+          0.5,
+          0.7142857142857143,
+          0.7142857142857143,
+          1.0,
+          1.0
+        ]
+      }
+    ]
+  },
+  "censoring_distribution": "weibull",
+  "dropout": 0.0,
+  "ensemble_size": 5,
+  "hidden_dim": 512,
+  "img_size": [
+    512,
+    512
+  ],
+  "initializer_range": 0.02,
+  "max_followup": 6,
+  "model_type": "sybil",
+  "num_images": 208,
+  "transformers_version": "4.53.2",
+  "voxel_spacing": [
+    0.703125,
+    0.703125,
+    2.5
+  ]
+}

configuration_sybil.py

@@ -0,0 +1,63 @@
+"""Sybil model configuration"""
+
+from transformers import PretrainedConfig
+from typing import Optional, List, Dict
+import json
+
+
+class SybilConfig(PretrainedConfig):
+    """
+    This is the configuration class to store the configuration of a [`SybilForRiskPrediction`].
+    It is used to instantiate a Sybil model according to the specified arguments, defining the model architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs.
+
+    Args:
+        hidden_dim (`int`, *optional*, defaults to 512):
+            Dimensionality of the hidden representations.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The dropout probability for all fully connected layers.
+        max_followup (`int`, *optional*, defaults to 6):
+            Maximum number of years for risk prediction.
+        num_images (`int`, *optional*, defaults to 208):
+            Number of CT scan slices to process.
+        img_size (`List[int]`, *optional*, defaults to `[512, 512]`):
+            Size of input images after preprocessing.
+        voxel_spacing (`List[float]`, *optional*, defaults to `[0.703125, 0.703125, 2.5]`):
+            Target voxel spacing for CT scans (row, column, slice thickness).
+        censoring_distribution (`str`, *optional*, defaults to "weibull"):
+            Distribution used for censoring in survival analysis.
+        ensemble_size (`int`, *optional*, defaults to 5):
+            Number of models in the ensemble.
+        calibrator_data (`Dict`, *optional*):
+            Calibration data for risk score adjustment.
+    """
+
+    model_type = "sybil"
+
+    def __init__(
+        self,
+        hidden_dim: int = 512,
+        dropout: float = 0.0,
+        max_followup: int = 6,
+        num_images: int = 208,
+        img_size: List[int] = None,
+        voxel_spacing: List[float] = None,
+        censoring_distribution: str = "weibull",
+        ensemble_size: int = 5,
+        calibrator_data: Optional[Dict] = None,
+        initializer_range: float = 0.02,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_dim = hidden_dim
+        self.dropout = dropout
+        self.max_followup = max_followup
+        self.num_images = num_images
+        self.img_size = img_size if img_size is not None else [512, 512]
+        self.voxel_spacing = voxel_spacing if voxel_spacing is not None else [0.703125, 0.703125, 2.5]
+        self.censoring_distribution = censoring_distribution
+        self.ensemble_size = ensemble_size
+        self.calibrator_data = calibrator_data
+        self.initializer_range = initializer_range

image_processing_sybil.py

@@ -0,0 +1,451 @@
+"""Image processor for Sybil CT scan preprocessing"""
+
+import cv2
+import numpy as np
+import torch
+from typing import Dict, List, Optional, Union, Tuple
+from transformers.image_processing_utils import BaseImageProcessor, BatchFeature
+from transformers.utils import TensorType
+import pydicom
+from PIL import Image
+import torchio as tio
+
+
+def order_slices(dicoms: List) -> List:
+    """Order DICOM slices by their position"""
+    # Sort by ImagePositionPatient if available
+    try:
+        dicoms = sorted(dicoms, key=lambda x: float(x.ImagePositionPatient[2]))
+    except (AttributeError, TypeError):
+        # Fall back to InstanceNumber if ImagePositionPatient not available
+        try:
+            dicoms = sorted(dicoms, key=lambda x: int(x.InstanceNumber))
+        except (AttributeError, TypeError):
+            pass  # Keep original order if neither attribute is available
+    return dicoms
+
+
+class SybilImageProcessor(BaseImageProcessor):
+    """
+    Constructs a Sybil image processor for preprocessing CT scans.
+
+    Args:
+        voxel_spacing (`List[float]`, *optional*, defaults to `[0.703125, 0.703125, 2.5]`):
+            Target voxel spacing for resampling (row, column, slice thickness).
+        img_size (`List[int]`, *optional*, defaults to `[512, 512]`):
+            Target image size after resizing.
+        num_images (`int`, *optional*, defaults to `208`):
+            Number of slices to use from the CT scan.
+        windowing (`Dict[str, float]`, *optional*):
+            Windowing parameters for CT scan visualization.
+            Default uses lung window: center=-600, width=1500.
+        normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize pixel values to [0, 1].
+        **kwargs:
+            Additional keyword arguments passed to the parent class.
+    """
+
+    model_input_names = ["pixel_values"]
+
+    def __init__(
+        self,
+        voxel_spacing: List[float] = None,
+        img_size: List[int] = None,
+        num_images: int = 208,
+        windowing: Dict[str, float] = None,
+        normalize: bool = True,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        self.voxel_spacing = voxel_spacing if voxel_spacing is not None else [0.703125, 0.703125, 2.5]
+        self.img_size = img_size if img_size is not None else [512, 512]
+        self.num_images = num_images
+
+        # Default lung window settings
+        self.windowing = windowing if windowing is not None else {
+            "center": -600,
+            "width": 1500
+        }
+        self.normalize = normalize
+
+        # TorchIO transforms for standardization
+        self.resample_transform = tio.transforms.Resample(target=self.voxel_spacing)
+        # Note: Original Sybil uses 200 depth, 256x256 images
+        self.default_depth = 200
+        self.default_size = [256, 256]
+        # TorchIO uses (H, W, D) ordering for target_shape, matching original Sybil
+        self.padding_transform = tio.transforms.CropOrPad(
+            target_shape=tuple(self.default_size + [self.default_depth]),  # (256, 256, 200)
+            padding_mode=0
+        )
+
+    def load_dicom_series(self, paths: List[str]) -> Tuple[np.ndarray, Dict]:
+        """
+        Load a series of DICOM files.
+
+        Args:
+            paths: List of paths to DICOM files.
+
+        Returns:
+            Tuple of (volume array, metadata dict)
+        """
+        dicoms = []
+        for path in paths:
+            try:
+                dcm = pydicom.dcmread(path, stop_before_pixels=False)
+                dicoms.append(dcm)
+            except Exception as e:
+                print(f"Error reading DICOM file {path}: {e}")
+                continue
+
+        if not dicoms:
+            raise ValueError("No valid DICOM files found")
+
+        # Order slices by position
+        dicoms = order_slices(dicoms)
+
+        # Extract pixel arrays
+        volume = np.stack([dcm.pixel_array.astype(np.float32) for dcm in dicoms])
+
+        # Extract metadata
+        metadata = {
+            "slice_thickness": float(dicoms[0].SliceThickness) if hasattr(dicoms[0], 'SliceThickness') else None,
+            "pixel_spacing": list(map(float, dicoms[0].PixelSpacing)) if hasattr(dicoms[0], 'PixelSpacing') else None,
+            "manufacturer": str(dicoms[0].Manufacturer) if hasattr(dicoms[0], 'Manufacturer') else None,
+            "num_slices": len(dicoms)
+        }
+
+        # Apply rescale if present
+        if hasattr(dicoms[0], 'RescaleSlope') and hasattr(dicoms[0], 'RescaleIntercept'):
+            slope = float(dicoms[0].RescaleSlope)
+            intercept = float(dicoms[0].RescaleIntercept)
+            volume = volume * slope + intercept
+
+        return volume, metadata
+
+    def load_png_series(self, paths: List[str]) -> np.ndarray:
+        """
+        Load a series of PNG files.
+
+        Args:
+            paths: List of paths to PNG files (must be in anatomical order).
+
+        Returns:
+            3D volume array
+        """
+        images = []
+        for path in paths:
+            img = Image.open(path).convert('L')  # Convert to grayscale
+            images.append(np.array(img, dtype=np.float32))
+
+        return np.stack(images)
+
+    def resize_slices(self, volume: np.ndarray, target_size: List[int] = None) -> np.ndarray:
+        """
+        Resize each slice in the volume to target size using OpenCV bilinear interpolation.
+        This exactly matches the original Sybil's per-slice 2D resize operation.
+
+        Args:
+            volume: 3D volume array (D, H, W).
+            target_size: Target size [H, W]. Defaults to [256, 256].
+
+        Returns:
+            Resized volume.
+        """
+        if target_size is None:
+            target_size = self.default_size  # [256, 256]
+
+        # Resize each slice using OpenCV (matching original Sybil exactly)
+        resized_slices = []
+        for i in range(volume.shape[0]):
+            slice_2d = volume[i]  # Shape: (H, W)
+            # cv2.resize expects dsize=(width, height), not (height, width)!
+            resized = cv2.resize(
+                slice_2d,
+                dsize=(target_size[1], target_size[0]),  # (W, H)
+                interpolation=cv2.INTER_LINEAR
+            )
+            resized_slices.append(resized)
+
+        # Stack back into volume
+        return np.stack(resized_slices, axis=0)
+
+    def apply_windowing(self, volume: np.ndarray) -> np.ndarray:
+        """
+        Apply DICOM-standard windowing to CT scan, matching the original Sybil implementation.
+
+        This implements the same windowing as the original Sybil:
+        - Uses DICOM standard formula with center-0.5 and width-1 adjustments
+        - Outputs to 16-bit range [0, 65535] then divides by 256 for 8-bit parity
+        - Results in [0, 255] range that will be normalized later
+
+        Args:
+            volume: 3D CT volume in Hounsfield Units.
+
+        Returns:
+            Windowed volume in [0, 255] range.
+        """
+        center = self.windowing["center"]  # -600
+        width = self.windowing["width"]     # 1500
+
+        # DICOM standard windowing formula (matching original Sybil)
+        bit_size = 16
+        y_min = 0
+        y_max = 2 ** bit_size - 1  # 65535
+        y_range = y_max - y_min
+
+        # DICOM standard adjustments
+        c = center - 0.5  # -600.5
+        w = width - 1     # 1499
+
+        # Calculate window boundaries
+        lower_bound = c - w / 2  # -1350
+        upper_bound = c + w / 2  # 149.5
+
+        # Apply windowing with three regions
+        below = volume <= lower_bound
+        above = volume > upper_bound
+        between = np.logical_and(~below, ~above)
+
+        # Create output array
+        windowed = np.zeros_like(volume, dtype=np.float32)
+
+        # Apply windowing
+        windowed[below] = y_min      # Values <= -1350 -> 0
+        windowed[above] = y_max      # Values > 149.5 -> 65535
+
+        if between.any():
+            # Linear interpolation for values in window
+            windowed[between] = ((volume[between] - c) / w + 0.5) * y_range + y_min
+
+        # Divide by 256 for 8-bit parity (matching original Sybil)
+        # This gives range [0, 255] instead of [0, 65535]
+        windowed = windowed // 256
+
+        return windowed
+
+    def resample_volume(
+        self,
+        volume: torch.Tensor,
+        original_spacing: Optional[List[float]] = None
+    ) -> torch.Tensor:
+        """
+        Resample volume to target voxel spacing.
+        Uses affine matrix approach matching original Sybil exactly.
+
+        Args:
+            volume: 3D or 4D volume tensor (D, H, W) or (C, D, H, W).
+            original_spacing: Original voxel spacing [H_spacing, W_spacing, D_spacing].
+
+        Returns:
+            Resampled volume with same number of dimensions.
+        """
+        # Handle both 3D (D, H, W) and 4D (C, D, H, W) volumes
+        if len(volume.shape) == 3:
+            # Single channel: (D, H, W) -> (1, D, H, W)
+            volume_4d = volume.unsqueeze(0)
+            squeeze_output = True
+        elif len(volume.shape) == 4:
+            # Multi-channel: (C, D, H, W) - already has channel dim
+            volume_4d = volume
+            squeeze_output = False
+        else:
+            raise ValueError(f"Expected 3D or 4D volume, got shape {volume.shape}")
+
+        # Permute to TorchIO format: (C, D, H, W) -> (C, H, W, D)
+        volume_tio = volume_4d.permute(0, 2, 3, 1)
+
+        # Create affine matrix like original Sybil
+        # Original uses torch.diag(voxel_spacing) where voxel_spacing has 4 elements
+        if original_spacing is not None:
+            # Add 1.0 as 4th element like original Sybil
+            voxel_spacing_4d = torch.tensor(original_spacing + [1.0], dtype=torch.float32)
+            affine = torch.diag(voxel_spacing_4d)
+        else:
+            affine = None
+
+        # Create TorchIO subject with affine (not spacing!)
+        subject = tio.Subject(
+            image=tio.ScalarImage(tensor=volume_tio, affine=affine)
+        )
+
+        # Apply resampling
+        resampled = self.resample_transform(subject)
+
+        # Permute back: (C, H, W, D) -> (C, D, H, W)
+        result = resampled['image'].data.permute(0, 3, 1, 2)
+
+        # Return with original number of dimensions
+        if squeeze_output:
+            return result.squeeze(0)
+        else:
+            return result
+
+    def pad_or_crop_volume(self, volume: torch.Tensor) -> torch.Tensor:
+        """
+        Pad or crop volume to target shape.
+
+        Args:
+            volume: 3D or 4D volume tensor (D, H, W) or (C, D, H, W).
+
+        Returns:
+            Padded/cropped volume with same number of dimensions.
+        """
+        # Handle both 3D (D, H, W) and 4D (C, D, H, W) volumes
+        if len(volume.shape) == 3:
+            # Single channel: (D, H, W) -> (1, D, H, W)
+            volume_4d = volume.unsqueeze(0)
+            squeeze_output = True
+        elif len(volume.shape) == 4:
+            # Multi-channel: (C, D, H, W) - already has channel dim
+            volume_4d = volume
+            squeeze_output = False
+        else:
+            raise ValueError(f"Expected 3D or 4D volume, got shape {volume.shape}")
+
+        # Permute to TorchIO format: (C, D, H, W) -> (C, H, W, D)
+        volume_tio = volume_4d.permute(0, 2, 3, 1)
+
+        # Create TorchIO subject
+        subject = tio.Subject(
+            image=tio.ScalarImage(tensor=volume_tio)
+        )
+
+        # Apply padding/cropping
+        transformed = self.padding_transform(subject)
+
+        # Permute back: (C, H, W, D) -> (C, D, H, W)
+        result = transformed['image'].data.permute(0, 3, 1, 2)
+
+        # Return with original number of dimensions
+        if squeeze_output:
+            return result.squeeze(0)
+        else:
+            return result
+
+    def preprocess(
+        self,
+        images: Union[List[str], np.ndarray, torch.Tensor],
+        file_type: str = "dicom",
+        voxel_spacing: Optional[List[float]] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        **kwargs
+    ) -> BatchFeature:
+        """
+        Preprocess CT scan images.
+
+        Args:
+            images: Either list of file paths or numpy/torch array of images.
+            file_type: Type of input files ("dicom" or "png").
+            voxel_spacing: Original voxel spacing (required for PNG files).
+            return_tensors: The type of tensors to return.
+
+        Returns:
+            BatchFeature with preprocessed images.
+        """
+        # Load images if paths are provided
+        if isinstance(images, list) and isinstance(images[0], str):
+            if file_type == "dicom":
+                volume, metadata = self.load_dicom_series(images)
+                if voxel_spacing is None and metadata["pixel_spacing"]:
+                    voxel_spacing = metadata["pixel_spacing"] + [metadata["slice_thickness"]]
+            elif file_type == "png":
+                if voxel_spacing is None:
+                    raise ValueError("voxel_spacing must be provided for PNG files")
+                volume = self.load_png_series(images)
+            else:
+                raise ValueError(f"Unknown file type: {file_type}")
+        elif isinstance(images, (np.ndarray, torch.Tensor)):
+            volume = images
+        else:
+            raise ValueError("Images must be file paths, numpy array, or torch tensor")
+
+        # Ensure volume is numpy array for initial processing
+        if isinstance(volume, torch.Tensor):
+            volume_np = volume.numpy()
+        else:
+            volume_np = volume
+
+        # Apply windowing
+        volume_np = self.apply_windowing(volume_np)
+
+        # Resize each slice to 256x256 (matching original Sybil's per-slice resize)
+        volume_np = self.resize_slices(volume_np, target_size=self.default_size)
+
+        # NOTE: Original Sybil uses the ORIGINAL voxel spacing from DICOM metadata
+        # even after resizing slices. This is physically incorrect (spacing should be
+        # adjusted for the resize factor), but we match the original behavior here.
+        # The voxel_spacing remains unchanged from DICOM metadata.
+
+        # Convert to torch tensor for remaining operations
+        volume = torch.from_numpy(volume_np).float()
+
+        # Apply normalization BEFORE resampling (to match original Sybil)
+        # Original Sybil normalizes each slice before assembly and 3D resampling
+        # This ensures 3D interpolation happens on normalized values, not [0, 255] values
+        # These values come from the original Sybil implementation's computed mean/std
+        # on 8-bit windowed images [0, 255]
+        img_mean = 128.1722
+        img_std = 87.1849
+        volume = (volume - img_mean) / img_std
+
+        # Replicate to 3 channels BEFORE resampling (to match original Sybil)
+        # Original Sybil replicates channels per-slice, then assembles 3-channel volume
+        # Shape: (D, H, W) -> (3, D, H, W)
+        volume = volume.unsqueeze(0).repeat(3, 1, 1, 1)  # Now (3, D, H, W)
+
+        # Resample if spacing is provided (3D resampling for voxel spacing adjustment)
+        # This happens on 3-channel volume, matching original Sybil
+        if voxel_spacing is not None:
+            volume = self.resample_volume(volume, voxel_spacing)
+
+        # Pad or crop to target shape (on 3-channel volume)
+        volume = self.pad_or_crop_volume(volume)
+
+        # Add batch dimension to match original Sybil output shape [1, C, D, H, W]
+        volume = volume.unsqueeze(0)  # Now (1, 3, D, H, W)
+
+        # Prepare output
+        data = {"pixel_values": volume}
+
+        # Convert to requested tensor type
+        if return_tensors == "pt":
+            return BatchFeature(data=data, tensor_type=TensorType.PYTORCH)
+        elif return_tensors == "np":
+            data = {k: v.numpy() for k, v in data.items()}
+            return BatchFeature(data=data, tensor_type=TensorType.NUMPY)
+        else:
+            return BatchFeature(data=data)
+
+    def __call__(
+        self,
+        images: Union[List[str], List[List[str]], np.ndarray, torch.Tensor],
+        **kwargs
+    ) -> BatchFeature:
+        """
+        Main method to prepare images for the model.
+
+        Args:
+            images: Images to preprocess. Can be:
+                - List of file paths for a single series
+                - List of lists of file paths for multiple series
+                - Numpy array or torch tensor
+
+        Returns:
+            BatchFeature with preprocessed images ready for model input.
+        """
+        # Handle batch processing
+        if isinstance(images, list) and images and isinstance(images[0], list):
+            # Multiple series
+            batch_volumes = []
+            for series_paths in images:
+                result = self.preprocess(series_paths, **kwargs)
+                batch_volumes.append(result["pixel_values"])
+
+            # Stack into batch (B, C, D, H, W)
+            pixel_values = torch.stack(batch_volumes)
+            return BatchFeature(data={"pixel_values": pixel_values})
+        else:
+            # Single series
+            return self.preprocess(images, **kwargs)

modeling_sybil.py

@@ -0,0 +1,603 @@
+"""PyTorch Sybil model for lung cancer risk prediction"""
+
+import torch
+import torch.nn as nn
+import torchvision
+from transformers import PreTrainedModel
+from transformers.modeling_outputs import BaseModelOutput
+from typing import Optional, Dict, List, Tuple
+import numpy as np
+from dataclasses import dataclass
+
+try:
+    from .configuration_sybil import SybilConfig
+except ImportError:
+    from configuration_sybil import SybilConfig
+
+
+@dataclass
+class SybilOutput(BaseModelOutput):
+    """
+    Base class for Sybil model outputs.
+
+    Args:
+        risk_scores: (`torch.FloatTensor` of shape `(batch_size, max_followup)`):
+            Predicted risk scores for each year up to max_followup.
+        image_attention: (`torch.FloatTensor` of shape `(batch_size, num_slices, height, width)`, *optional*):
+            Attention weights over image pixels.
+        volume_attention: (`torch.FloatTensor` of shape `(batch_size, num_slices)`, *optional*):
+            Attention weights over CT scan slices.
+        hidden_states: (`torch.FloatTensor` of shape `(batch_size, hidden_dim)`, *optional*):
+            Hidden states from the pooling layer.
+    """
+    risk_scores: torch.FloatTensor = None
+    image_attention: Optional[torch.FloatTensor] = None
+    volume_attention: Optional[torch.FloatTensor] = None
+    hidden_states: Optional[torch.FloatTensor] = None
+
+
+class CumulativeProbabilityLayer(nn.Module):
+    """
+    Cumulative probability layer for survival prediction.
+
+    Matches the original Sybil implementation exactly with:
+    - hazard_fc: Year-specific hazards (can be zero after ReLU)
+    - base_hazard_fc: Base hazard shared across all years
+    - Triangular masking for cumulative hazard computation
+    """
+
+    def __init__(self, hidden_dim: int, max_followup: int = 6):
+        super().__init__()
+        self.max_followup = max_followup
+
+        # Year-specific hazards
+        self.hazard_fc = nn.Linear(hidden_dim, max_followup)
+
+        # Base hazard (shared across years)
+        self.base_hazard_fc = nn.Linear(hidden_dim, 1)
+
+        self.relu = nn.ReLU(inplace=True)
+
+        # Upper triangular mask for cumulative computation
+        mask = torch.ones([max_followup, max_followup])
+        mask = torch.tril(mask, diagonal=0)
+        mask = torch.nn.Parameter(torch.t(mask), requires_grad=False)
+        self.register_parameter("upper_triangular_mask", mask)
+
+    def hazards(self, x):
+        """Compute positive hazards using ReLU"""
+        raw_hazard = self.hazard_fc(x)
+        pos_hazard = self.relu(raw_hazard)
+        return pos_hazard
+
+    def forward(self, x):
+        """
+        Compute cumulative probabilities matching original Sybil.
+
+        Args:
+            x: Hidden features [B, hidden_dim]
+
+        Returns:
+            Cumulative probabilities [B, max_followup]
+        """
+        hazards = self.hazards(x)
+        B, T = hazards.size()
+
+        # Expand for masking: [B, T] -> [B, T, T]
+        expanded_hazards = hazards.unsqueeze(-1).expand(B, T, T)
+
+        # Apply triangular mask for cumulative sum
+        masked_hazards = expanded_hazards * self.upper_triangular_mask
+
+        # Base hazard (shared across years)
+        base_hazard = self.base_hazard_fc(x)
+
+        # Sum masked hazards and add base
+        cum_prob = torch.sum(masked_hazards, dim=1) + base_hazard
+
+        return cum_prob
+
+
+class GlobalMaxPool(nn.Module):
+    """Pool to obtain the maximum value for each channel"""
+
+    def __init__(self):
+        super(GlobalMaxPool, self).__init__()
+
+    def forward(self, x):
+        """
+        Args:
+            - x: tensor of shape (B, C, T, W, H)
+        Returns:
+            - output: dict. output['hidden'] is (B, C)
+        """
+        spatially_flat_size = (*x.size()[:2], -1)
+        x = x.view(spatially_flat_size)
+        hidden, _ = torch.max(x, dim=-1)
+        return {'hidden': hidden}
+
+
+class PerFrameMaxPool(nn.Module):
+    """Pool to obtain the maximum value for each slice in 3D input"""
+
+    def __init__(self):
+        super(PerFrameMaxPool, self).__init__()
+
+    def forward(self, x):
+        """
+        Args:
+            - x: tensor of shape (B, C, T, W, H)
+        Returns:
+            - output: dict.
+                + output['multi_image_hidden'] is (B, C, T)
+        """
+        assert len(x.shape) == 5
+        output = {}
+        spatially_flat_size = (*x.size()[:3], -1)
+        x = x.view(spatially_flat_size)
+        output['multi_image_hidden'], _ = torch.max(x, dim=-1)
+        return output
+
+
+class Simple_AttentionPool(nn.Module):
+    """Pool to learn an attention over the slices"""
+
+    def __init__(self, **kwargs):
+        super(Simple_AttentionPool, self).__init__()
+        self.attention_fc = nn.Linear(kwargs['num_chan'], 1)
+        self.softmax = nn.Softmax(dim=-1)
+        self.logsoftmax = nn.LogSoftmax(dim=-1)
+
+    def forward(self, x):
+        """
+        Args:
+            - x: tensor of shape (B, C, N)
+        Returns:
+            - output: dict
+                + output['volume_attention']: tensor (B, N)
+                + output['hidden']: tensor (B, C)
+        """
+        output = {}
+        B = x.shape[0]
+        spatially_flat_size = (*x.size()[:2], -1)  # B, C, N
+
+        x = x.view(spatially_flat_size)
+        attention_scores = self.attention_fc(x.transpose(1, 2))  # B, N, 1
+
+        output['volume_attention'] = self.logsoftmax(attention_scores.transpose(1, 2)).view(B, -1)
+        attention_scores = self.softmax(attention_scores.transpose(1, 2))  # B, 1, N
+
+        x = x * attention_scores  # B, C, N
+        output['hidden'] = torch.sum(x, dim=-1)
+        return output
+
+
+class Simple_AttentionPool_MultiImg(nn.Module):
+    """Pool to learn an attention over the slices and the volume"""
+
+    def __init__(self, **kwargs):
+        super(Simple_AttentionPool_MultiImg, self).__init__()
+        self.attention_fc = nn.Linear(kwargs['num_chan'], 1)
+        self.softmax = nn.Softmax(dim=-1)
+        self.logsoftmax = nn.LogSoftmax(dim=-1)
+
+    def forward(self, x):
+        """
+        Args:
+            - x: tensor of shape (B, C, T, W, H)
+        Returns:
+            - output: dict
+                + output['image_attention']: tensor (B, T, W*H)
+                + output['multi_image_hidden']: tensor (B, C, T)
+                + output['hidden']: tensor (B, T*C)
+        """
+        output = {}
+        B, C, T, W, H = x.size()
+        x = x.permute([0, 2, 1, 3, 4])
+        x = x.contiguous().view(B*T, C, W*H)
+        attention_scores = self.attention_fc(x.transpose(1, 2))  # BT, WH, 1
+
+        output['image_attention'] = self.logsoftmax(attention_scores.transpose(1, 2)).view(B, T, -1)
+        attention_scores = self.softmax(attention_scores.transpose(1, 2))  # BT, 1, WH
+
+        x = x * attention_scores  # BT, C, WH
+        x = torch.sum(x, dim=-1)
+        output['multi_image_hidden'] = x.view(B, T, C).permute([0, 2, 1]).contiguous()
+        output['hidden'] = x.view(B, T * C)
+        return output
+
+
+class Conv1d_AttnPool(nn.Module):
+    """Pool to learn an attention over the slices after convolution"""
+
+    def __init__(self, **kwargs):
+        super(Conv1d_AttnPool, self).__init__()
+        self.conv1d = nn.Conv1d(
+            kwargs['num_chan'],
+            kwargs['num_chan'],
+            kernel_size=kwargs['conv_pool_kernel_size'],
+            stride=kwargs['stride'],
+            padding=kwargs['conv_pool_kernel_size']//2,
+            bias=False
+        )
+        self.aggregate = Simple_AttentionPool(**kwargs)
+
+    def forward(self, x):
+        """
+        Args:
+            - x: tensor of shape (B, C, T)
+        Returns:
+            - output: dict
+                + output['attention_scores']: tensor (B, C)
+                + output['hidden']: tensor (B, C)
+        """
+        # X: B, C, N
+        x = self.conv1d(x)  # B, C, N'
+        return self.aggregate(x)
+
+
+class MultiAttentionPool(nn.Module):
+    """Multi-attention pooling layer for CT scan aggregation - matches original Sybil architecture"""
+
+    def __init__(self, channels: int = 512):
+        super().__init__()
+        params = {
+            'num_chan': 512,
+            'conv_pool_kernel_size': 11,
+            'stride': 1
+        }
+
+        # Define all pooling sub-modules matching original Sybil
+        self.image_pool1 = Simple_AttentionPool_MultiImg(**params)
+        self.volume_pool1 = Simple_AttentionPool(**params)
+        self.image_pool2 = PerFrameMaxPool()
+        self.volume_pool2 = Conv1d_AttnPool(**params)
+        self.global_max_pool = GlobalMaxPool()
+
+        # Final linear layers to combine features
+        self.multi_img_hidden_fc = nn.Linear(2 * 512, 512)
+        self.hidden_fc = nn.Linear(3 * 512, 512)
+
+    def forward(self, x):
+        """
+        Args:
+            x: tensor of shape (B, C, T, W, H) where
+               - B: batch size
+               - C: channels (512)
+               - T: temporal/depth dimension (slices)
+               - W, H: spatial dimensions
+
+        Returns:
+            output: dict with keys:
+                - 'hidden': (B, 512) - final aggregated features
+                - 'image_attention_1': (B, T, W*H) - image attention scores
+                - 'volume_attention_1': (B, T) - volume attention scores
+                - 'image_attention_2': None (no attention for max pool)
+                - 'volume_attention_2': (B, T) - volume attention scores
+                - 'multi_image_hidden': (B, 512, T) - intermediate features
+                - 'maxpool_hidden': (B, 512) - max pooled features
+        """
+        output = {}
+
+        # First attention pooling pathway
+        image_pool_out1 = self.image_pool1(x)
+        # Keys: "multi_image_hidden" (B, C, T), "image_attention" (B, T, W*H), "hidden" (B, T*C)
+
+        volume_pool_out1 = self.volume_pool1(image_pool_out1['multi_image_hidden'])
+        # Keys: "hidden" (B, C), "volume_attention" (B, T)
+
+        # Second max pooling pathway
+        image_pool_out2 = self.image_pool2(x)
+        # Keys: "multi_image_hidden" (B, C, T)
+
+        volume_pool_out2 = self.volume_pool2(image_pool_out2['multi_image_hidden'])
+        # Keys: "hidden" (B, C), "volume_attention" (B, T)
+
+        # Collect all pooling outputs with numbered suffixes
+        for pool_out, num in [(image_pool_out1, 1), (volume_pool_out1, 1),
+                               (image_pool_out2, 2), (volume_pool_out2, 2)]:
+            for key, val in pool_out.items():
+                output['{}_{}'.format(key, num)] = val
+
+        # Global max pooling
+        maxpool_out = self.global_max_pool(x)
+        output['maxpool_hidden'] = maxpool_out['hidden']
+
+        # Combine multi-image features from both pathways
+        multi_image_hidden = torch.cat(
+            [image_pool_out1['multi_image_hidden'], image_pool_out2['multi_image_hidden']],
+            dim=-2
+        )  # (B, C, 2*T)
+        output['multi_image_hidden'] = self.multi_img_hidden_fc(
+            multi_image_hidden.permute([0, 2, 1]).contiguous()
+        ).permute([0, 2, 1]).contiguous()  # (B, 512, T)
+
+        # Combine all volume-level features
+        hidden = torch.cat(
+            [volume_pool_out1['hidden'], volume_pool_out2['hidden'], output['maxpool_hidden']],
+            dim=-1
+        )  # (B, 3*512)
+        output['hidden'] = self.hidden_fc(hidden)  # (B, 512)
+
+        return output
+
+
+class SybilPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface
+    for downloading and loading pretrained models.
+    """
+    config_class = SybilConfig
+    base_model_prefix = "sybil"
+    supports_gradient_checkpointing = False
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Conv3d):
+            nn.init.kaiming_normal_(module.weight, mode='fan_out', nonlinearity='relu')
+            if module.bias is not None:
+                module.bias.data.zero_()
+
+
+class SybilForRiskPrediction(SybilPreTrainedModel):
+    """
+    Sybil model for lung cancer risk prediction from CT scans.
+
+    This model takes 3D CT scan volumes as input and predicts cancer risk scores
+    for multiple future time points (typically 1-6 years).
+    """
+
+    def __init__(self, config: SybilConfig):
+        super().__init__(config)
+        self.config = config
+
+        # Use pretrained R3D-18 as backbone
+        encoder = torchvision.models.video.r3d_18(pretrained=True)
+        self.image_encoder = nn.Sequential(*list(encoder.children())[:-2])
+
+        # Multi-attention pooling
+        self.pool = MultiAttentionPool(channels=512)
+
+        # Classification layers
+        self.relu = nn.ReLU(inplace=False)
+        self.dropout = nn.Dropout(p=config.dropout)
+
+        # Risk prediction layer
+        self.prob_of_failure_layer = CumulativeProbabilityLayer(
+            config.hidden_dim,
+            max_followup=config.max_followup
+        )
+
+        # Calibrator for ensemble predictions
+        self.calibrator = None
+        if config.calibrator_data:
+            self.set_calibrator(config.calibrator_data)
+
+        # Initialize weights
+        self.post_init()
+
+    def set_calibrator(self, calibrator_data: Dict):
+        """Set calibration data for risk score adjustment"""
+        self.calibrator = calibrator_data
+
+    def _calibrate_scores(self, scores: torch.Tensor) -> torch.Tensor:
+        """Apply calibration to raw risk scores"""
+        if self.calibrator is None:
+            return scores
+
+        # Convert to numpy for calibration
+        scores_np = scores.detach().cpu().numpy()
+        calibrated = np.zeros_like(scores_np)
+
+        # Apply calibration for each year
+        for year in range(scores_np.shape[1]):
+            year_key = f"Year{year + 1}"
+            if year_key in self.calibrator:
+                # Apply calibration transformation
+                calibrated[:, year] = self._apply_calibration(
+                    scores_np[:, year],
+                    self.calibrator[year_key]
+                )
+            else:
+                calibrated[:, year] = scores_np[:, year]
+
+        return torch.from_numpy(calibrated).to(scores.device)
+
+    def _apply_calibration(self, scores: np.ndarray, calibrator_params: Dict) -> np.ndarray:
+        """Apply specific calibration transformation"""
+        # Simplified calibration - in practice, this would use the full calibration model
+        # from the original Sybil implementation
+        return scores  # Placeholder for now
+
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        return_attentions: bool = False,
+        return_dict: bool = True,
+    ) -> SybilOutput:
+        """
+        Forward pass of the Sybil model.
+
+        Args:
+            pixel_values: (`torch.FloatTensor` of shape `(batch_size, channels, depth, height, width)`):
+                Pixel values of CT scan volumes.
+            return_attentions: (`bool`, *optional*, defaults to `False`):
+                Whether to return attention weights.
+            return_dict: (`bool`, *optional*, defaults to `True`):
+                Whether to return a `SybilOutput` instead of a plain tuple.
+
+        Returns:
+            `SybilOutput` or tuple
+        """
+        # Extract features using 3D CNN backbone
+        features = self.image_encoder(pixel_values)
+
+        # Apply multi-attention pooling
+        pool_output = self.pool(features)
+
+        # Apply ReLU and dropout
+        hidden = self.relu(pool_output['hidden'])
+        hidden = self.dropout(hidden)
+
+        # Predict risk scores
+        risk_logits = self.prob_of_failure_layer(hidden)
+        risk_scores = torch.sigmoid(risk_logits)
+
+        # Apply calibration if available
+        risk_scores = self._calibrate_scores(risk_scores)
+
+        if not return_dict:
+            outputs = (risk_scores,)
+            if return_attentions:
+                outputs = outputs + (pool_output.get('image_attention_1'),
+                                   pool_output.get('volume_attention_1'))
+            return outputs
+
+        return SybilOutput(
+            risk_scores=risk_scores,
+            image_attention=pool_output.get('image_attention_1') if return_attentions else None,
+            volume_attention=pool_output.get('volume_attention_1') if return_attentions else None,
+            hidden_states=hidden if return_attentions else None
+        )
+
+    @classmethod
+    def from_pretrained_ensemble(
+        cls,
+        pretrained_model_name_or_path,
+        checkpoint_paths: List[str],
+        calibrator_path: Optional[str] = None,
+        **kwargs
+    ):
+        """
+        Load an ensemble of Sybil models from checkpoints.
+
+        Args:
+            pretrained_model_name_or_path: Path to the pretrained model or model identifier.
+            checkpoint_paths: List of paths to individual model checkpoints.
+            calibrator_path: Path to calibration data.
+            **kwargs: Additional keyword arguments for model initialization.
+
+        Returns:
+            SybilEnsemble: An ensemble of Sybil models.
+        """
+        config = kwargs.pop("config", None)
+        if config is None:
+            config = SybilConfig.from_pretrained(pretrained_model_name_or_path)
+
+        # Load calibrator if provided
+        calibrator_data = None
+        if calibrator_path:
+            import json
+            with open(calibrator_path, 'r') as f:
+                calibrator_data = json.load(f)
+            config.calibrator_data = calibrator_data
+
+        # Create ensemble
+        models = []
+        for checkpoint_path in checkpoint_paths:
+            model = cls(config)
+            # Load checkpoint weights
+            checkpoint = torch.load(checkpoint_path, map_location='cpu')
+            # Remove 'model.' prefix from state dict keys if present
+            state_dict = {}
+            for k, v in checkpoint['state_dict'].items():
+                if k.startswith('model.'):
+                    state_dict[k[6:]] = v
+                else:
+                    state_dict[k] = v
+
+            # Map to new model structure
+            mapped_state_dict = model._map_checkpoint_weights(state_dict)
+            model.load_state_dict(mapped_state_dict, strict=False)
+            models.append(model)
+
+        return SybilEnsemble(models, config)
+
+    def _map_checkpoint_weights(self, state_dict: Dict) -> Dict:
+        """Map original Sybil checkpoint weights to new structure"""
+        mapped = {}
+
+        # Map encoder weights
+        for k, v in state_dict.items():
+            if k.startswith('image_encoder'):
+                mapped[k] = v
+            elif k.startswith('pool'):
+                # Map pooling layer weights
+                mapped[k] = v
+            elif k.startswith('prob_of_failure_layer'):
+                # Map final prediction layer
+                mapped[k] = v
+
+        return mapped
+
+
+class SybilEnsemble:
+    """Ensemble of Sybil models for improved predictions"""
+
+    def __init__(self, models: List[SybilForRiskPrediction], config: SybilConfig):
+        self.models = models
+        self.config = config
+        self.device = None
+
+    def to(self, device):
+        """Move all models to device"""
+        self.device = device
+        for model in self.models:
+            model.to(device)
+        return self
+
+    def eval(self):
+        """Set all models to evaluation mode"""
+        for model in self.models:
+            model.eval()
+
+    def __call__(
+        self,
+        pixel_values: torch.FloatTensor,
+        return_attentions: bool = False,
+    ) -> SybilOutput:
+        """
+        Run inference with ensemble voting.
+
+        Args:
+            pixel_values: Input CT scan volumes.
+            return_attentions: Whether to return attention maps.
+
+        Returns:
+            SybilOutput with averaged predictions from all models.
+        """
+        all_risk_scores = []
+        all_image_attentions = []
+        all_volume_attentions = []
+
+        with torch.no_grad():
+            for model in self.models:
+                output = model(
+                    pixel_values=pixel_values,
+                    return_attentions=return_attentions
+                )
+                all_risk_scores.append(output.risk_scores)
+
+                if return_attentions:
+                    all_image_attentions.append(output.image_attention)
+                    all_volume_attentions.append(output.volume_attention)
+
+        # Average predictions
+        risk_scores = torch.stack(all_risk_scores).mean(dim=0)
+
+        # Average attentions if requested
+        image_attention = None
+        volume_attention = None
+        if return_attentions:
+            image_attention = torch.stack(all_image_attentions).mean(dim=0)
+            volume_attention = torch.stack(all_volume_attentions).mean(dim=0)
+
+        return SybilOutput(
+            risk_scores=risk_scores,
+            image_attention=image_attention,
+            volume_attention=volume_attention
+        )

modeling_sybil_hf.py

@@ -0,0 +1,346 @@
+"""
+Self-contained Hugging Face wrapper for Sybil lung cancer risk prediction model.
+This version works directly from HF without requiring external Sybil package.
+"""
+
+import os
+import json
+import sys
+import torch
+import numpy as np
+from typing import List, Dict, Optional
+from dataclasses import dataclass
+from transformers.modeling_outputs import BaseModelOutput
+from safetensors.torch import load_file
+
+# Add model path to sys.path for imports
+current_dir = os.path.dirname(os.path.abspath(__file__))
+if current_dir not in sys.path:
+    sys.path.insert(0, current_dir)
+
+try:
+    from .configuration_sybil import SybilConfig
+    from .modeling_sybil import SybilForRiskPrediction
+    from .image_processing_sybil import SybilImageProcessor
+except ImportError:
+    from configuration_sybil import SybilConfig
+    from modeling_sybil import SybilForRiskPrediction
+    from image_processing_sybil import SybilImageProcessor
+
+
+@dataclass
+class SybilOutput(BaseModelOutput):
+    """
+    Output class for Sybil model predictions.
+
+    Args:
+        risk_scores: Risk scores for each year (1-6 years by default)
+        attentions: Optional attention maps if requested
+    """
+    risk_scores: torch.FloatTensor = None
+    attentions: Optional[Dict] = None
+
+
+class SybilHFWrapper:
+    """
+    Hugging Face wrapper for Sybil ensemble model.
+    Provides a simple interface for lung cancer risk prediction from CT scans.
+    """
+
+    def __init__(self, config: SybilConfig = None, model_dir: str = None):
+        """
+        Initialize the Sybil model ensemble.
+
+        Args:
+            config: Model configuration (will use default if not provided)
+            model_dir: Directory containing model files (defaults to file location)
+        """
+        self.config = config if config is not None else SybilConfig()
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+        # Get the directory where model files are located
+        if model_dir is not None:
+            self.model_dir = model_dir
+        else:
+            # Default to where this file is located
+            self.model_dir = os.path.dirname(os.path.abspath(__file__))
+
+        # Initialize image processor
+        self.image_processor = SybilImageProcessor()
+
+        # Load calibrator
+        self.calibrator = self._load_calibrator()
+
+        # Load ensemble models
+        self.models = self._load_ensemble_models()
+
+    def _load_calibrator(self) -> Dict:
+        """Load ensemble calibrator data"""
+        calibrator_path = os.path.join(self.model_dir, "checkpoints", "sybil_ensemble_simple_calibrator.json")
+
+        if os.path.exists(calibrator_path):
+            with open(calibrator_path, 'r') as f:
+                return json.load(f)
+        else:
+            # Try alternative location
+            calibrator_path = os.path.join(self.model_dir, "calibrator_data.json")
+            if os.path.exists(calibrator_path):
+                with open(calibrator_path, 'r') as f:
+                    return json.load(f)
+        return {}
+
+    def _load_ensemble_models(self) -> List[torch.nn.Module]:
+        """
+        Load all models in the ensemble from original checkpoints.
+
+        Note: We load from .ckpt files instead of safetensors because the safetensors
+        were created with the wrong CumulativeProbabilityLayer architecture.
+        """
+        import glob as glob_module
+        models = []
+
+        # Find all .ckpt files in checkpoints directory
+        checkpoints_dir = os.path.join(self.model_dir, "checkpoints")
+        checkpoint_files = sorted(glob_module.glob(os.path.join(checkpoints_dir, "*.ckpt")))
+
+        print(f"Found {len(checkpoint_files)} checkpoint files")
+
+        # Load each checkpoint file
+        for checkpoint_path in checkpoint_files:
+            try:
+                model = SybilForRiskPrediction(self.config)
+                checkpoint = torch.load(checkpoint_path, map_location='cpu', weights_only=False)
+
+                # Extract state dict
+                if 'state_dict' in checkpoint:
+                    state_dict = checkpoint['state_dict']
+                else:
+                    state_dict = checkpoint
+
+                # Remove 'model.' prefix if present
+                cleaned_state_dict = {}
+                for k, v in state_dict.items():
+                    if k.startswith('model.'):
+                        cleaned_state_dict[k[6:]] = v
+                    else:
+                        cleaned_state_dict[k] = v
+
+                # Load weights
+                model.load_state_dict(cleaned_state_dict, strict=False)
+                model.to(self.device)
+                model.eval()
+                models.append(model)
+                print(f"  Loaded model from {os.path.basename(checkpoint_path)}")
+            except Exception as e:
+                print(f"  Warning: Could not load {os.path.basename(checkpoint_path)}: {e}")
+                continue
+
+        if not models:
+            raise ValueError("No models could be loaded from the ensemble. Please ensure model files are present.")
+
+        print(f"Loaded {len(models)} models in ensemble")
+        return models
+
+    def _apply_calibration(self, scores: np.ndarray) -> np.ndarray:
+        """
+        Apply complete isotonic regression calibration matching the original Sybil implementation.
+
+        This method applies the same calibration as the original SimpleClassifierGroup.predict_proba:
+        1. For each year, apply each calibrator in the ensemble
+        2. Each calibrator applies: linear transform -> clip -> isotonic regression (np.interp)
+        3. Average predictions from all calibrators
+
+        Args:
+            scores: Raw risk scores from the model (shape: [batch_size, num_years])
+
+        Returns:
+            Calibrated risk scores (shape: [batch_size, num_years])
+        """
+        if not self.calibrator:
+            return scores
+
+        calibrated_scores = []
+
+        for year in range(scores.shape[1]):
+            year_key = f"Year{year + 1}"
+
+            if year_key not in self.calibrator:
+                # No calibrator for this year, use raw scores
+                calibrated_scores.append(scores[:, year])
+                continue
+
+            cal_list = self.calibrator[year_key]
+
+            if not isinstance(cal_list, list) or len(cal_list) == 0:
+                # Invalid calibrator format, use raw scores
+                calibrated_scores.append(scores[:, year])
+                continue
+
+            # Apply each calibrator and collect predictions
+            year_predictions = []
+
+            for cal_data in cal_list:
+                if not isinstance(cal_data, dict):
+                    continue
+
+                # Extract calibration parameters
+                if "coef" not in cal_data or "intercept" not in cal_data:
+                    continue
+
+                coef = np.array(cal_data["coef"])  # Shape: [[scalar]]
+                intercept = np.array(cal_data["intercept"])  # Shape: [scalar]
+
+                # Extract isotonic regression points
+                if "x0" not in cal_data or "y0" not in cal_data:
+                    continue
+
+                x0 = np.array(cal_data["x0"])
+                y0 = np.array(cal_data["y0"])
+
+                # Extract clipping bounds
+                x_min = cal_data.get("x_min", -np.inf)
+                x_max = cal_data.get("x_max", np.inf)
+
+                # Apply complete calibration pipeline:
+                # Step 1: Linear transformation
+                probs = scores[:, year].reshape(-1, 1)  # Shape: [batch_size, 1]
+                T = probs @ coef + intercept  # Matrix multiplication
+                T = T.flatten()  # Shape: [batch_size]
+
+                # Step 2: Clip to valid range
+                T = np.clip(T, x_min, x_max)
+
+                # Step 3: Apply isotonic regression via interpolation
+                # This is the CRITICAL step that was missing!
+                calibrated = np.interp(T, x0, y0)
+
+                year_predictions.append(calibrated)
+
+            if len(year_predictions) == 0:
+                # No valid calibrators, use raw scores
+                calibrated_scores.append(scores[:, year])
+            else:
+                # Average predictions from all calibrators (like SimpleClassifierGroup)
+                calibrated_scores.append(np.mean(year_predictions, axis=0))
+
+        return np.stack(calibrated_scores, axis=1)
+
+    def preprocess_dicom(self, dicom_paths: List[str]) -> torch.Tensor:
+        """
+        Preprocess DICOM files for model input.
+
+        Args:
+            dicom_paths: List of paths to DICOM files
+
+        Returns:
+            Preprocessed tensor ready for model input
+        """
+        # Use the image processor to handle DICOM files
+        result = self.image_processor(dicom_paths, file_type="dicom", return_tensors="pt")
+        pixel_values = result["pixel_values"]
+
+        # Ensure we have 5D tensor (B, C, D, H, W)
+        if pixel_values.ndim == 4:
+            pixel_values = pixel_values.unsqueeze(0)  # Add batch dimension
+
+        return pixel_values.to(self.device)
+
+    def predict(self, dicom_paths: List[str], return_attentions: bool = False) -> SybilOutput:
+        """
+        Run prediction on a CT scan series.
+
+        Args:
+            dicom_paths: List of paths to DICOM files for a single CT series
+            return_attentions: Whether to return attention maps
+
+        Returns:
+            SybilOutput with risk scores and optional attention maps
+        """
+        # Preprocess the DICOM files
+        pixel_values = self.preprocess_dicom(dicom_paths)
+
+        # Run inference with ensemble
+        all_predictions = []
+        all_attentions = []
+
+        with torch.no_grad():
+            for model in self.models:
+                output = model(
+                    pixel_values=pixel_values,
+                    return_attentions=return_attentions
+                )
+
+                # Extract risk scores
+                if hasattr(output, 'risk_scores'):
+                    predictions = output.risk_scores
+                else:
+                    predictions = output[0] if isinstance(output, tuple) else output
+
+                all_predictions.append(predictions.cpu().numpy())
+
+                if return_attentions and hasattr(output, 'image_attention'):
+                    all_attentions.append(output.image_attention)
+
+        # Average ensemble predictions
+        ensemble_pred = np.mean(all_predictions, axis=0)
+
+        # Apply calibration
+        calibrated_pred = self._apply_calibration(ensemble_pred)
+
+        # Convert back to torch tensor
+        risk_scores = torch.from_numpy(calibrated_pred).float()
+
+        # Average attentions if requested
+        attentions = None
+        if return_attentions and all_attentions:
+            attentions = {"image_attention": torch.stack(all_attentions).mean(dim=0)}
+
+        return SybilOutput(risk_scores=risk_scores, attentions=attentions)
+
+    def __call__(self, dicom_paths: List[str] = None, dicom_series: List[List[str]] = None, **kwargs) -> SybilOutput:
+        """
+        Convenience method for prediction.
+
+        Args:
+            dicom_paths: List of DICOM file paths for a single series
+            dicom_series: List of lists of DICOM paths for batch processing
+            **kwargs: Additional arguments passed to predict()
+
+        Returns:
+            SybilOutput with predictions
+        """
+        if dicom_series is not None:
+            # Batch processing
+            all_outputs = []
+            for paths in dicom_series:
+                output = self.predict(paths, **kwargs)
+                all_outputs.append(output.risk_scores)
+
+            risk_scores = torch.stack(all_outputs)
+            return SybilOutput(risk_scores=risk_scores)
+        elif dicom_paths is not None:
+            return self.predict(dicom_paths, **kwargs)
+        else:
+            raise ValueError("Either dicom_paths or dicom_series must be provided")
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
+        """
+        Load model from Hugging Face hub or local path.
+
+        Args:
+            pretrained_model_name_or_path: HF model ID or local path
+            **kwargs: Additional configuration arguments
+
+        Returns:
+            SybilHFWrapper instance
+        """
+        # Load configuration
+        config = kwargs.pop("config", None)
+        if config is None:
+            try:
+                config = SybilConfig.from_pretrained(pretrained_model_name_or_path)
+            except:
+                config = SybilConfig()
+
+        return cls(config=config)

requirements.txt

@@ -0,0 +1,9 @@
+torch>=2.0.0
+torchvision>=0.15.0
+transformers>=4.30.0
+pydicom>=2.3.0
+torchio>=0.19.0
+SimpleITK>=2.2.0
+numpy>=1.21.0
+Pillow>=9.0.0
+sybil>=1.2.0