Automated optic chiasm segmentation in magnetic resonance imaging of multiple sclerosis patients using a multi-view, multi-class U-Net with limited annotated data

Background: Accurate magnetic resonance imaging (MRI)-based segmentation of the optic chiasm is essential for the quantitative assessment of visual pathway damage in multiple sclerosis (MS). However, accurate segmentation remains challenging because of the structure’s small size, complex anatomy, and marked foreground–background imbalance. Existing deep learning approaches often rely on single-view imaging and may not fully exploit complementary anatomical information across MRI planes. To address these limitations, we developed and evaluated a novel multi-view two-dimensional U-Net (2D U-Net) framework that jointly leverages axial and coronal MRI information within a unified segmentation architecture for automated optic chiasm segmentation. Materials and Methods: In this secondary analysis, a novel multi-view 2D U-Net framework was developed for optic chiasm segmentation by integrating axial and coronal MRI images within a unified three-class segmentation architecture. Models were optimized using a Tversky loss function and evaluated using five-fold patient-wise cross-validation. Segmentation performance was assessed quantitatively using multiple metrics, with the Dice similarity coefficient (DSC), intersection over union (IoU), and balanced accuracy serving as the primary evaluation metrics, and qualitatively through visual comparison with expert annotations. Results: A total of 58 patients with relapsing-remitting MS were included in the analysis. The proposed multi-view model demonstrated stable convergence and achieved the highest segmentation performance among all evaluated models, with a DSC of 91.28%, IoU of 84.05%, and balanced accuracy of 94.43%. Compared with the best-performing single-view model, the multi-view approach improved DSC and IoU by approximately 13 and 19 percentage points, respectively. Qualitative assessment showed closer agreement between multi-view segmentations and expert annotations. Conclusion: The proposed multi-view, multi-class 2D U-Net framework outperformed conventional single-view models for optic chiasm segmentation. These findings demonstrate that integrating complementary anatomical information from axial and coronal MRI planes can substantially improve segmentation performance. The results indicate the potential of multi-view learning for the automated segmentation of small and anatomically complex structures.