A CT foundation model built for cancer patients' chest-abdomen-pelvis scans lifts AUC to 0.767 from 0.695

Chest-abdomen-pelvis CT in cancer patients now has a foundation model trained for that clinical setting, and it leads all compared models on disease classification in the full-coverage and oncologic subsets. Earlier CT foundation models either covered a single anatomical region, such as chest only or abdomen-pelvis only, or covered broader anatomy without oncology-rich training data. On the chest-abdomen-pelvis subset CAPTOR reached an AUC of 0.767 (95% CI 0.734–0.793) against 0.695 (0.645–0.736) for the next-best model; on the oncologic subset it reached 0.784 (0.759–0.805) against 0.734 (0.703–0.767), both p<0.001.

Rather than compressing a whole scan into one image, the model first cuts the exam into organ and anatomical-region crops and encodes each separately, so one architecture handles full chest-abdomen-pelvis studies. Earlier region-specific models used fixed central crops, which struggle to cover the full chest-abdomen-pelvis extent and multi-series acquisitions. Crop centers come from TotalSegmentator masks; each training batch rotates through one anatomical region and shuffles between epochs so every crop of a study is seen, with pretraining data from 11 institutions in North America, Europe and Asia.

The results cover study-level understanding: the model aggregates large anatomical regions into a single study-level embedding for disease classification, two-year survival prediction and image retrieval, and is not designed for lesion localization or characterization. Evaluation used an internal multicenter test set and public datasets, with classification stratified by anatomical coverage and oncologic versus non-oncologic indication. Code and model weights are released for reproducing training and inference.

The study is retrospective with no prospective validation. Reference labels were extracted automatically from reports by a large language model, with extraction quality assessed on a public dataset and 25 randomly selected reports manually reviewed, so omissions in the original reports and model errors may still introduce label noise at scale. The four external oncologic cohorts are small, limiting statistical power for pairwise comparisons. The model was outperformed by generalist models on non-oncologic chest CT datasets, indicating that specialization for a target setting may cost performance on substantially different imaging populations.