chn123/astra-encoder-nlst-features

🫁 ASTRA-Sybil & ASTRA-Pillar: NLST 3D CT Survival Benchmark & Feature Archive

This repository hosts pre-extracted 3D spatial feature tensors and the canonical biostatistical validation benchmark for deep lung cancer risk assessment on the National Lung Screening Trial (NLST).

---

📁 Repository Structure

.
├── 5D/                                         # 2,965 PyTorch 5D spatial feature tensors [2048, 10, 7, 7]
│   └── *.pt
├── manifest.csv                                # Full dataset manifest with clinical survival labels
├── reports/
│   ├── ASTRA_SYBIL_PILLAR_PROGRESS_REPORT.docx # Comprehensive scientific & clinical report (Word DOCX)
│   └── ASTRA_SYBIL_PILLAR_PROGRESS_REPORT.md   # Markdown progress report with full ablation tables
├── canonical_validation/                       # Canonical Biostatistical Validation Framework
│   ├── canonical_metrics.py                    # Harrell C, Uno IPCW C, t-AUC, Brier, Calibration, Bootstrap
│   ├── run_canonical_reevaluation.py           # Patient-level evaluation of V1, V2, V3 (1000 bootstraps)
│   ├── run_baselines_e4.py                     # Survival Baseline Ladder (KM, Cox, Astra 1D Linear Probe)
│   ├── run_ablations_e1_e2_e3.py               # GPU-Resident Ablation Suite (Pooling, Loss, Stratification)
│   └── results/
│       ├── canonical_v1_v2_v3_benchmark.json   # Full metric JSON with 95% CIs and paired p-values
│       ├── e4_survival_baselines_benchmark.json# Survival ladder JSON
│       └── e1_e2_e3_ablations_benchmark.json   # E1, E2, E3 ablations JSON
├── scripts/                                    # Training and extraction scripts
│   ├── train_sybil_vs_pillar_astra_v3.py       # V3 Model (Joint Temporal Stratification + Ranking Loss)
│   ├── train_sybil_vs_pillar_astra_v2.py       # V2 Model (Patient-Level Stratification)
│   └── train_sybil_vs_pillar_astra.py          # V1 Model (Series GroupKFold)
└── benchmarks/                                 # OOF predictions, metrics, and attention maps
    ├── sybil_oof_survival_v3_predictions.csv
    ├── pillar_oof_survival_v3_predictions.csv
    └── positive_cases_attention_maps.npz

---

🔬 Clinical Cohort Overview

• Dataset: National Lung Screening Trial (NLST) Low-Dose CT (LDCT)
• Unique Patients ($N$): 1,201 (Zero patient overlap across folds)
• Total Scans / Series: 2,965
• Positive (Incident Cancer) Cases: 72 patients (140 series; 6.0% incidence)
• Censored / Cancer-Free Controls: 1,129 patients (2,825 series)
• Visual Encoder: Astra / Merlin (Inflated 3D ResNet-152, I3ResNet)
• Feature Tensor Dimensions: [2048, 10, 7, 7] (490 spatial tokens per scan)

---

📊 Canonical Survival Benchmark Results (Patient-Level, $N=1,201$)

All evaluations are conducted at the independent patient level ($N=1,201$) with maximum risk aggregation across serial scans. Confidence intervals [95% CI] and paired $p$-values are computed via 1,000 cluster bootstrap resamples.

1. Survival Baseline Ladder (E4)

Baseline Model	Harrell's C-Index [95% CI]	Uno's IPCW C-Index	Year 1 AUC [95% CI]	Year 6 AUC [95% CI]	Integrated Brier Score (IBS)
Kaplan-Meier (Unconditional)	0.4924 [0.4263 - 0.5620]	0.5000	0.4171 [0.3340 - 0.5050]	0.5000 [0.5000 - 0.5000]	0.0700
Clinical Cox Model (Age, Gender, Smoking)	0.6446 [0.5786 - 0.7101]	0.6585	0.5993 [0.5008 - 0.6970]	0.6486 [0.5762 - 0.7188]	0.0685
Astra 1D Linear Probe (GAP + Ridge)	0.6764 [0.6108 - 0.7409]	0.7185	0.6540 [0.5606 - 0.7441]	0.6877 [0.6200 - 0.7513]	0.0641
Clinical + Astra 1D Combined	0.6893 [0.6262 - 0.7500]	0.7310	0.6594 [0.5663 - 0.7508]	0.7033 [0.6385 - 0.7645]	0.0636
Best Deep 3D Head (Astra-Pillar V3)	0.6241 [0.5625 - 0.6824]	0.6295	0.5689 [0.4687 - 0.6657]	0.6648 [0.5960 - 0.7288]	0.0716

Key Scientific Insight: Astra visual representations carry a genuine, strong prognostic signal (pushing concordance from 0.64 to 0.69). However, because the cohort contains only 72 incident events, heavily regularized linear probes on Astra GAP features outperform 3.1M-parameter deep spatial attention heads ($p \gg N_{events}$).

---

2. Pooling Mechanism Ablation (E1)

Pooling Architecture	Trainable Parameters	Harrell's C-Index [95% CI]	Year 1 AUC [95% CI]	Integrated Brier Score (IBS)
Global MaxPool (GMP)	1,053,703	0.5745 [0.5056 - 0.6428]	0.5977 [0.5112 - 0.6817]	0.0674
Global AvgPool (GAP)	1,053,703	0.6159 [0.5454 - 0.6839]	0.5579 [0.4673 - 0.6422]	0.0668
Spatial Gated MIL	2,103,048	0.6060 [0.5364 - 0.6715]	0.5667 [0.4655 - 0.6632]	0.0669
Astra-Sybil Pooling	2,630,665	0.6423 [0.5768 - 0.7103]	0.5795 [0.4826 - 0.6767]	0.0721
Astra-Pillar Pooling	3,154,952	0.6121 [0.5462 - 0.6771]	0.5873 [0.4896 - 0.6810]	0.0716

Differences between Sybil and Pillar are not statistically significant ($p = 0.440$). Global AvgPool achieves significantly better Brier score calibration ($p = 0.010$).

---

3. Loss Formulation Ablation (E2)

Loss Objective	Harrell C [95% CI]	Uno IPCW C	Year 1 AUC [95% CI]	IBS	Calibration Assessment
Masked BCE	0.5851 [0.5183 - 0.6506]	0.5682	0.5914 [0.4914 - 0.6900]	0.0811	Suboptimal Brier score
BCE + Pairwise Ranking	0.6191 [0.5566 - 0.6857]	0.6143	0.6447 [0.5484 - 0.7322]	0.0690	Best discrimination & calibration
Pure Pairwise Ranking	0.6091 [0.5367 - 0.6794]	0.6592	0.5381 [0.4375 - 0.6329]	0.1153	Uncalibrated (Poor Brier error)
Discrete Hazard NLL	0.5249 [0.4568 - 0.5951]	0.5319	0.5818 [0.4879 - 0.6637]	0.2499	Gradient sparsity collapse

---

4. Gradient-Free Spatial Probing Benchmark (E5: Testing NN Optimization Failure)

Operator	Type	Dimension	Harrell C-Index [95% CI]	Uno IPCW C	Year 1 AUC	Paired ΔC vs GAP [95% CI]
Top-2 Pool (~0.4% vol)	Local Nodule Scale	2048	0.6832 [0.6126 - 0.7501]	0.6942	0.6840	+0.0059 [-0.0340, +0.0452]
GAP + Spatial Std	Hybrid (Global+Spread)	4096	0.6812 [0.6143 - 0.7459]	0.7082	0.6804	+0.0047 [-0.0191, +0.0269]
Top-1 GMP (Global Max)	Peak Activation	2048	0.6796 [0.6074 - 0.7480]	0.7042	0.6658	+0.0022 [-0.0405, +0.0437]
Spatial Std (Heterogeneity)	Dispersion Moment	2048	0.6789 [0.6097 - 0.7434]	0.7010	0.6881	+0.0024 [-0.0237, +0.0286]
GAP Baseline	Global Mean	2048	0.6769 [0.6151 - 0.7389]	0.7129	0.6422	0.0000 [REFERENCE]

---

5. Controlled Focal-Spatial & Multimodal Clinical Fusion Benchmark (E6)

Evaluated under strict 5-Fold Stratified Cross-Validation at the Patient Level ($N=1,201$) with 3-fold nested inner-CV for hyperparameter tuning.

Model ID	Feature Set	Harrell C [95% CI]	Uno C	AUC 1y	AUC 2y	AUC 3y	AUC 4y	AUC 5y
M0	Clinical Demographics Only	0.6435 [0.5748 - 0.7110]	0.6411	0.5817	0.6167	0.6294	0.6417	0.6456
M1	GAP Imaging Alone	0.6278 [0.5577 - 0.6939]	0.6439	0.5766	0.6708	0.6001	0.6098	0.6273
M2	Clinical + GAP (Reference)	0.6258 [0.5562 - 0.6924]	0.6426	0.5760	0.5989	0.5818	0.6120	0.6245
M3	Top-2 Imaging Alone	0.6775 [0.6076 - 0.7491]	0.6961	0.6267	0.6777	0.6687	0.6910	0.6815
M4	Clinical + Top-2 (Primary)	0.6783 [0.6080 - 0.7501]	0.6968	0.6260	0.6778	0.6682	0.6912	0.6823
M6	Clinical + Spatial Std	0.6607 [0.5917 - 0.7294]	0.6777	0.6213	0.6118	0.6610	0.6817	0.6747
M8	Clinical + GAP + Std	0.6462 [0.5801 - 0.7159]	0.6640	0.5988	0.6021	0.6234	0.6533	0.6545
M10	Clinical + Top-2 + Std	0.6655 [0.5968 - 0.7376]	0.6824	0.6255	0.6058	0.6473	0.6791	0.6735

🎯 Primary Endpoint Hypothesis Test:

$$\mathrm{\Delta }C=C_{extClinical+Top-2}-C_{extClinical+GAP}=+\mathbf{0.0529}[95\mathrm{\%}extCI:+\mathbf{0.0058},+\mathbf{0.1018}],p=\mathbf{0.0260}(extSTATISTICALLYSIGNIFICANT)\backslash Delta\; C\; =\; C\_\{\; ext\{Clinical\; +\; Top-2\}\}\; -\; C\_\{\; ext\{Clinical\; +\; GAP\}\}\; =\; \backslash mathbf\{+0.0529\}\; \backslash \; [95\backslash \%\; ext\{\; CI:\; \}\; \backslash mathbf\{+0.0058,\; \backslash \; +0.1018\}],\; \backslash \; p\; =\; \backslash mathbf\{0.0260\}\; \backslash \; (\; ext\{STATISTICALLY\; SIGNIFICANT\})$$

📉 PCA Dimensionality Compression (8 to 32 Dimensions Reaching C ≈ 0.72):

Representation	Raw 2048/4096D	PCA-8	PCA-16	PCA-32	PCA-64	PCA-128
Clinical + Top-2	0.6783	0.7188	0.6895	0.6993	0.6739	0.6548
Clinical + Top-2 + Std	0.6655	0.7203	0.6895	0.6921	0.6765	0.6330
Clinical + Spatial Std	0.6607	0.7186	0.6888	0.6920	0.6627	0.6140
Clinical + GAP	0.6258	0.6889	0.6702	0.6664	0.6720	0.6072

---

🚀 Loading 5D Spatial Tensors

import torch
from huggingface_hub import hf_hub_download

# Download a sample 5D tensor
path = hf_hub_download(
    repo_id="chn123/astra-encoder-nlst-features",
    subfolder="5D",
    filename="100012_01-02-1999-NLST-LSS-56831_2.000000-0OPASEVZOOMB30f3002.012075.040.0null-00079.pt",
    repo_type="dataset"
)

tensor = torch.load(path, weights_only=True)
print("Shape:", tensor.shape)  # torch.Size([2048, 10, 7, 7])

---

📖 Citation & References

• Astra: Multimodal CT Foundation Model, Baykar / Astra Team, 2024.
• Sybil: Mikhael et al., Journal of Clinical Oncology, 2023.
• Pillar: Yala et al., Nature Medicine / Lancet Digital Health, 2024.