This repository is a reproducible research pipeline for an approximately 1B-parameter CT oncology foundation model. It is not a trained clinical checkpoint and is not approved for diagnosis, triage, or treatment selection. Training and clinical validation require approved data, qualified investigators, and execution on GPUs.

The pipeline uses a deliberate dataset portfolio rather than pretending that one public dataset contains all required labels.

The dataset rationale and primary source URLs are recorded in dataset_catalog.yaml and oncology_ct_research_notes.md.

The repository never silently downloads an entire archive or bypasses access controls. Acquire only approved data and retain citations and license metadata.

For IDC, generate a cohort using idc-index or IDC BigQuery, export the query result, and keep the per-file license in each manifest row. For TCIA imaging, use the official manifest with TCIA Data Retriever after accepting the applicable terms. Do not mix patient identifiers across source systems without an approved linkage protocol.

The preprocessing contract is NIfTI-first: CT volumes are clipped to [-1024, 3071] HU, normalized, patchified into 8³ voxel tokens, and saved as .pt tensors with metadata. DICOM series and DICOM-SEG/RTSTRUCT should be converted using an audited SimpleITK/dcmqi pipeline before invoking the NIfTI preprocessor.

Build and validate a unified manifest as follows:

Every row must include patient_id, institution, tensor_path, source_dataset, license, split, and label provenance. Splits are patient- and institution-scoped; any patient appearing in multiple splits is a hard error. Treatment and follow-up labels must be timestamped so future information never leaks into baseline prediction.

The default model uses 1,536-dimensional embeddings, 32 Transformer blocks, 24 attention heads, 8³ voxel patches, and up to 4,096 volume tokens. It exposes lesion, response, quality-control, volume, long-axis, embedding, and abstention outputs. Optional report embeddings can be supplied, but the product boundary remains evidence extraction and clinician review rather than autonomous treatment recommendation.

Training proceeds from broad CT representation learning on IDC/MSD/ULS23, to lesion characterization, to longitudinal correspondence on Longitudinal-CT, to treatment-response multitask learning on HCC-TACE-Seg, followed by institution-held-out evaluation. NSCLC-Radiomics is retained as a noncommercial research evaluation source unless its license is separately cleared.

Read deploy/README.md and replace placeholders only after confirming the current Google Cloud regional accelerator catalog, quota, machine family, networking, and storage policies. Use FSDP or the supported distributed launcher for the 50-GPU job. Start with one GPU, then 2–8 GPUs, then a controlled multi-node run with checkpoint recovery before using the full quota.

The delivered source is research-only. Any clinical pilot requires institutional review, data-use agreements, de-identification verification, prospective protocol, external site validation, calibration and subgroup analysis, human-factor testing, audit logs, model-version pinning, drift monitoring, rollback, and a defined regulatory intended use. The service must abstain when input quality, domain shift, or model confidence is inadequate.