A 3D SAM-Based Progressive Prompting Framework for Multi-Task Segmentation of Radiotherapy-induced Normal Tissue Injuries in Limited-Data Settings
Summary: arXiv:2604.13367v1 Announce Type: cross
Abstract
Radiotherapy-induced normal tissue injury is a clinically significant complication that necessitates accurate segmentation of injury regions from medical images. Such segmentation can greatly facilitate disease assessment, treatment planning, and longitudinal monitoring. Despite its importance, the automatic segmentation of these lesions remains largely unexplored due to limited voxel-level annotations and substantial heterogeneity across various injury types, lesion sizes, and imaging modalities.
Introduction
To address the challenges associated with the segmentation of radiotherapy-induced injuries, researchers have curated a dedicated dataset focusing on head-and-neck radiotherapy-induced normal tissue injuries. This dataset encompasses three primary manifestations:
- Osteoradionecrosis (ORN)
- Cerebral edema (CE)
- Cerebral radiation necrosis (CRN)
Each of these conditions presents unique challenges for segmentation due to variations in their characteristics and the complexity of the images involved.
The Proposed Framework
In response to the limited-data settings, the researchers propose a novel 3D SAM-based progressive prompting framework designed for multi-task segmentation. This framework effectively integrates three complementary prompts:
- Text Prompts: These are used for task-aware adaptation, allowing the model to understand the specific requirements of each segmentation task.
- Dose-Guided Box Prompts: These prompts enable coarse localization of injuries, providing the model with initial guidance on the potential locations of lesions.
- Click Prompts: These facilitate iterative refinement, allowing for adjustments based on user input to improve the accuracy of segmentation.
Innovation in Loss Function
A significant innovation in this framework is the introduction of a small-target focus loss function. This loss function is specifically designed to enhance local predictions and improve boundary delineation for small and sparse lesions, which are often difficult to segment accurately.
Experimental Results
Comprehensive experiments were conducted on the dataset encompassing ORN, CE, and CRN. The results demonstrate that the proposed method achieves reliable segmentation performance across diverse injury types. Remarkably, it outperforms several state-of-the-art methods, highlighting the efficacy of the three-prompt strategy and the innovative loss function in achieving accurate segmentation in challenging scenarios.
Conclusion
The advancements in segmentation techniques for radiotherapy-induced normal tissue injuries represent a significant step forward in medical imaging. The 3D SAM-based progressive prompting framework not only addresses the limitations posed by limited data but also sets a new benchmark for future research in this critical area of healthcare. By improving the accuracy and reliability of injury segmentation, this framework could substantially enhance patient outcomes through better assessment and treatment planning.