Multi-contrast MR-driven deep learning for abdominal multi-organ segmentation (McDAMOS)

Motivation: Efficient and accurate contouring of abdominal organs-at-risk (OAR) is crucial for MR-guided radiotherapy planning and online adaptation but challenging due to complex anatomy. Multi-contrast MR may be utilized to achieve automated multi-organ segmentation. Goal(s): To develop a multi-contrast MR-driven DL technique for abdominal multi-organ segmentation. Approach: Our model builds on a 3D Swin Transformer architecture with T1w and T2w dual inputs. Pre-training on a larger T1w dataset and synthesized T2w images addressed limited data. A VAE-based loss for organ shape learning was incorporated. Results: Multi-contrast inputs, pre-training, and VAE loss all contributed to improved segmentation performance, especially for challenging organs like the duodenum. Impact: Our work demonstrates the utility of multi-contrast MR in achieving abdominal auto-segmentation and presents a methodology to address limited data available from a novel research MR sequence. The approach benefits clinicians and propelling automated segmentation techniques forward.