2591 - Supporting High-Precision Stereotactic Radiosurgery: A Multi-Institution Evaluation of a Deep Learning Substructure Model for Brain Organ at Risk Segmentation in MRI

Purpose/Objective(s): Stereotactic radiosurgery (SRS) delivers very high radiation doses with steep dose gradients, thus accurate organ at risk (OAR) delineation is essential. Recent evidence suggests that dose to brain substructures may relate to critical neurologic functions. To support consistent and precise delineation, we developed a deep learning brain substructure model and evaluated its performance on multi-institutional datasets to ensure robustness across scanners, imaging protocols (without (T1) or with (T1C) contrast), and patient populations.

Materials/Methods: A set of 60 cases (3D Ax T1 BRAVO) were collected at Institution 1, 50 (35 T1C) were used for training, 10 (6 T1C) for testing. The mean FOV (250 mm), slice coverage (170 mm), and pixel size (0.5 mm) were fixed for all scans, but the slice thickness was smaller (0.6/1.2 mm) for T1C scans. Another set of 10 (4 T1C) test cases (3D Ax T1 MP-RAGE) were collected at Institution 2. The mean FOV (256 mm), slice coverage (256 mm), and slice thickness (1 mm) were fixed for all scans, but the pixel size was smaller (0.5/1 mm) for T1C scans. Ground truth contours for 28 organs and substructures including deep gray matter nuclei, and ventricular system were delineated based on radiation therapy guidelines. Using the 50 annotated cases, a multi-class, 3D nnU-Net was trained, and evaluated with DICE similarity. Two-sample T-test was applied to show significant difference.

Results: Table 1 presents the overall and organ specific mean DICE measured in 4 subsets of the total 20 test cases. The results show that the model performed equally on T1 and T1C scans, however, it had significantly (p<0.01) lower accuracy on independent (Institution 2) test data, though the overall difference was modest (84.8% vs. 87.6%). Longer scan extent contributed to the uncertainty in Institution 2 as result of which the model overestimated mainly the brain and the cerebellum in the inferior part that was not covered during training with Institution 1 scans.

Conclusion: Deep learning-based brain segmentation can achieve high accuracy across standard MR imaging parameters and contrast conditions. Modest performance reduction observed on independent test data highlights the need for expanding training data diversity to further improve generalizability and support more reliable OAR delineation in SRS.