Mayo Clinic's AI Spots Radiation Injuries Doctors Struggle to Track
A new 3D SAM framework layers text, dose-map, and click prompts to segment tiny radiotherapy injuries from just 70 patient scans.
Takeaways
- Mayo Clinic team builds a 3D SAM framework for segmenting radiotherapy-induced tissue injuries.
- Handles three injury types (ORN, cerebral edema, radiation necrosis) in one unified model.
- Uses progressive prompts: text for task, dose-map box for localization, clicks for refinement.
- Trained on just 70 curated head-and-neck cases, split roughly 4:1.
- Hits 77.11 Dice and cuts HD95 from 9.80 to 5.70 versus SwinUNETR.
- Small-target focus loss restricts optimization to the clinically relevant high-dose ROI.
Radiation therapy saves lives, but the same beams that kill tumor cells can leave behind collateral damage in nearby healthy tissue. That damage, called radiotherapy-induced normal tissue injury, is hard to track over time because the lesions are tiny, sparsely scattered, and rarely annotated at the voxel level. A team at Mayo Clinic's Department of Radiation Oncology has now built a segmentation framework that handles three different injury types with a single model, using fewer than a hundred patient cases.
The paper targets three specific complications in head-and-neck cancer patients: osteoradionecrosis (ORN), cerebral edema (CE), and cerebral radiation necrosis (CRN). Each shows up on a different imaging modality, each has a different shape, and none of them have large public datasets. That combination has kept the problem mostly untouched by modern deep learning.
Why standard segmentation networks fall apart here
Three things make radiotherapy injury segmentation unusually painful. Automatic delineation of these lesions remains largely unexplored because of limited voxel-level annotations and substantial heterogeneity across injury types, lesion size, and imaging modality. Injury regions are also typically extremely small relative to the surrounding volume, which means a naive voxel-wise loss gets swamped by background. And because the same patient can develop different injury types either together or sequentially, training one model per manifestation is wasteful and misses the shared structure across tasks.
Existing work in this area has focused mostly on predicting which patients are at risk rather than drawing the actual lesion outlines. The handful of segmentation attempts that do exist are usually specialized to a single injury in a single scenario, using off-the-shelf networks like Mask R-CNN or DeepMedic.