Isarar/Evidence-Field

Evidence Field — source-selected research checkpoints

Convolity AI Lab: Isarar Siddique, Raiyan Siddique, Arshi Niyaz, Jamyang Ugyen Tshomo and Atish Raza.

Implementation and independent worked-example reproduction: isararsiddique/evidence-field.

What is supplied

This repository supplies six original source-selected EF state dictionaries: three CAMELYON17 histopathology members and three known lung-nodule CT members, with seeds 11/29/47. The two tasks have distinct checkpoints/adapters and are not a single universally validated medical classifier. The ensemble configuration files contain the original architecture parameters, source-selection records, class labels, positive temperatures and weights. Each member has 10,550 registered parameters.

Evidence_Field_Model_Package.zip is self-contained: exact numerical code, CT descriptor/direct-execution helpers, six original checkpoints, source locks, protocols, configurations, a hash manifest and load_models.py. Root-level .pt files are the same weights, provided for direct downloads. The loader uses torch.load(..., weights_only=True) and strict state loading; no training runs. No dataset images, credentials or source-validation prediction arrays are supplied.

Load with the original implementation

Download and extract Evidence_Field_Model_Package.zip, install Python 3.12, torch==2.7.0, NumPy and SciPy, then run in separate fresh processes:

python EF_Models/load_models.py --task histopathology --device cpu
python EF_Models/load_models.py --task lung --device cpu

from EF_Models.load_models import load_ensemble
model = load_ensemble('histopathology', device='cpu')
# prepared_images: N x 3 x 96 x 96 float32 RGB /255
result = model.predict(prepared_images, calibrated=False)

Use calibrated=True only when the recorded source temperatures are intended. Probabilities are averaged after each member's softmax/temperature; logits are not averaged. The return contains member logits/probabilities, ensemble probabilities and predicted class labels. Native signed-program export and framework-free reconstruction are documented in the linked code repository. This custom architecture is loaded with the supplied implementation, not an automatic Transformers AutoModel or hosted inference pipeline.

Required input contracts

• Histopathology: float32 N x 3 x 96 x 96, native lymph-node RGB divided by 255. Released label: tumour in the central 32 x 32 window. This is not whole-slide/patient diagnosis or colorectal tissue classification.
• Lung CT: float32 N x 1 x 64 x 64 x 64, known-nodule raw-HU crops under the registered resampling/cropping protocol supplied inside Lung_EF/. This is nodule classification, not whole-scan detection. Do not substitute RGB, arbitrary intensity normalization or a different crop protocol.

Evaluation and limitations

Primary histopathology raw-ensemble AUROC/accuracy is 0.9356 / 84.46% at CAMELYON17 centre 1 and 0.9453 / 86.77% at centre 2. Source training comes from centres 0/3/4; source validation alone selected checkpoints/objectives. These are two held-out hospitals within one dataset. They had been examined in earlier work: source-locked replication is not a prospective trial or the standard WILDS leaderboard protocol. The code repository records exact patient partitions, mirror provenance, paired patient-bootstrap uncertainty and the unequal tuning budget of from-scratch CNN/ViT comparisons.

Trained Gaussian-closure/fixed-partition controls outperform primary EF at both hospitals on AUROC point estimates. Learned signatures/adaptive partitions therefore do not have established predictive benefit in this study. The signed covering/replay contract is distinct from diagnostic superiority or causal localization. Lung configurations retain their own source-validation metrics; the histopathology metrics above must not be attributed to the CT models.

Models and evidence outputs are research artifacts. General 2D/3D mathematics does not establish universal radiology accuracy, clinical safety, independent doctor review, certification or publication acceptance. Third-party datasets retain their original access/licensing conditions.

Complete runnable histopathology application

Download Histopathology_EF_Ready_To_Run.zip. It includes all three original histopathology weights, the complete unchanged encoder/classifier/recursion, file-based RGB preprocessing, prediction CLI, independently reconstructed evidence outputs, three real examples including an error, and a tested dependency/usage guide. After extraction, run python predict_histopathology.py --input examples/example_negative.png --output results/negative. The trained contract is a native 96x96 lymph-node RGB patch; it is not a whole-slide or universal pathology classifier.

Fresh CPU inference reproduced the recorded decisions on 340 real histology audit patches and 85 real CT crops (all 73 LUNGx plus 12 LIDC examples), 1,275 member-image predictions. The downloaded original Spaces archive has the same six checkpoint hashes. See functional verification and its machine-readable receipt. This is prediction reproduction, not new training or an expanded diagnostic accuracy claim.