Multi-Anatomy Geometric Accuracy Study for MRI Based Deep Learning Segmentation In MR-Only Radiotherapy Planning
Abstract
Purpose
MR-only radiotherapy (RT) planning faces workflow challenges that limits its use in routine practice. Currently, clinics rely on manual contouring and/or deep learning based automatic segmentation (DLAS) trained using clinical-specific MRI sequences. The former is time consuming; the latter is inefficient for practices where MRI sequences evolve as the needs of the clinic evolve. We evaluate here the feasibility of using vendor provided DLAS models for male and female pelvis (MP, FP) and head and neck (HN) disease sites.
Methods
Segmentation of 50 structures was performed with a DLAS model trained using a ResUNet3D network and 209 multi-institutional MR-linac MRIs from the MOMENTUM clinical trial (NCT04075303). The models were implemented in a research tool (ADMIRE v4.2, Elekta) and were evaluated using 39 independent MR simulation and MR-Linac datasets from our Institution. Segmentation accuracy was evaluated visually slice-by-slice for HN and using 6 geometric accuracy metrics including DSC with ground truth (GT) contours for MP and FP sites.
Results
Segmentation time was <30 sec for all datasets. The MP segmentation outperformed across three unique MRI sequences (DSCAvg 0.72-0.97). 85% contour clinical acceptability was observed for HN segmentation. Poor segmentation of rectum in MP was attributed to inter-observer variability at sigmoid- and anus-rectum junctions and improved accuracy was seen when length matched between DLAS and GT (17% for MP and 30% for FP). DLAS accuracy was found to be almost independent of MRI pre-processing technique. Observed limitations included sequence dependence, inaccurate bowel segmentation (DSC 0.58), incomplete segmentation of small/long thin structures e.g., optic nerves in HN) and inaccurate labeling of gauze packing/ring/ovoid applicator as rectum in FP.
Conclusion
These results demonstrate that MR-based DLAS models trained using non-clinic-specific MRI sequences can generate reasonable baseline contours for HN and near perfect baseline contours in MP to support routine clinical MR-only RT planning.