A deep learning model that segments 24 abdominal organs from a CT volume, with an early experimental head for per organ cancer risk. Initialized from SuPreM and fine tuned on the AbdomenAtlas dataset.
This project segments 24 abdominal structures from a single CT volume. On top of that segmentation sits an early, experimental head that estimates per organ cancer risk, a research direction rather than a diagnostic. Locating early, localized disease depends first on knowing precisely where each organ sits, a problem studied by the Johns Hopkins BodyMaps group.
A 3D SwinUNETR backbone is initialized from the SuPreM pretrained weights and fine tuned on AbdomenAtlas. Transfer learning does most of the work when labeled data is limited.
Rare, thin structures such as vessels, the adrenal glands and the celiac trunk are sampled deliberately rather than left to chance. This raised mean Dice from 0.19 to 0.62 during training.
An experimental per organ risk head is trained with synthetic tumor augmentation. It has not been validated on real tumor scans and is a research direction, not a diagnostic. Monte Carlo dropout maps show where the model is least certain, pointing to regions a radiologist may want to review.
Gaussian weighted sliding window inference reconstructs a segmentation of the whole scan. Scores are measured on full volumes, the stricter approach used by benchmarks such as Touchstone.
The full 24 organ breakdown, including the harder thin structures such as the celiac trunk, hepatic vessel and femurs, is available in the technical report.
The model was trained on 199 of the 9,262 available cases. Full volume mean Dice is 0.62, against 0.85 and higher reported at full scale. That gap reflects the volume of training data rather than the design of the model, and narrows as the data scales. The cancer risk head is experimental and has not been validated on real tumor scans.
The trained model and evaluation code can be shared for academic and research purposes. To request access or discuss a collaboration, reach out on LinkedIn.