Automated Segmentation of the Brainstem Using Deep Learning U-Net Architectures Applied to Multiplanar Magnetic Resonance Images

Neuroimaging research has identified MRI biomarkers that may aid the detection and staging of Parkinson's disease (PD), including neuromelanin signals in the substantia nigra and locus coeruleus.Current analyses rely on costly manual segmentation.AI models like U-Nets achieve strong results in medical imaging but are rarely applied to brainstem structures due to limited datasets.In this paper, we trained U-Nets on a T1-weighted MRI dataset developed by our group, starting with Freesurfer-based brainstem masks validated by specialists.Four models were evaluated using the Dice Similarity Coefficient (DSC), the Intersection over Union (IoU), Hausdorff Distance (HD) and some of its variations.The coronal model performed best (DSC 95.34%, IoU 93.03%, HD 0.68), followed by axial (93.88%, 89.17%, 1.89, 0.09, 0.70) and the multi-plane approach (92.42%, 87.49%, 1.95).Our future work will refine midbrain segmentation, apply new thresholding, and analyze hemispheres to advance PD biomarker detection.