Cohort-Aware CT Reproduction

This repository implements the first-stage reproduction scaffold for paper.pdf: CT-first pulmonary lesion characterization with optional PET, clinical missingness, segmentation support maps, calibration hooks and discrete-time survival output.

Current Scope

Implemented: LIDC/Lung-PET-CT-Dx/NSCLC manifest loaders, DICOM/NIfTI/NPZ preprocessing hooks, 3D support network, CT/PET encoders, selective PET gate, clinical encoder, survival head, losses, metrics, training and evaluation CLI.
Reserved: NLST schema and survival interface only. The raw NLST data is not downloaded in this phase.
Deferred: exact TCIA RTSTRUCT/XML parsing details and pathology regularization until the corresponding data is present locally.

Environment

Use the requested conda environment:

conda activate ct
python -m pip install -r requirements-project.txt

Smoke Test

conda run -n ct python scripts/create_synthetic_manifest.py --out-dir processed/synthetic --n 8
conda run -n ct pytest -q

Training Entry Points

conda run -n ct python -m src.train --config configs/lidc_seg.yaml
conda run -n ct python -m src.train --config configs/lung_pet_ct_dx.yaml
conda run -n ct python -m src.train --config configs/nsclc_external.yaml

Manifests must point to preprocessed lesion crops. Each row should include ct_path; optional columns are pet_path, mask_path, label, time, event, patient_id, lesion_id and configured clinical columns.