Substantia Nigra Imaging Biomarker Segmentation for Parkinson's Disease Diagnosis via Transformer-Enhanced U-Net Architecture.
Authors
Abstract
The substantia nigra pars compacta (SNpc) provides critical imaging biomarkers for Parkinson's disease (PD) when accurately segmented from magnetic resonance imaging scans. The extremely limited spatial extent of SNpc regions, combined with subtle tissue-intensity variations and ambiguous boundaries shared with adjacent midbrain structures, poses major challenges for existing segmentation methods that rely solely on local feature extraction. To enhance computer-aided diagnosis through improved SNpc segmentation, we propose SNIB-TransNet, a substantia nigra imaging biomarker transformer network built on an enhanced U-shaped architecture. The method combines multi-head attention with skip connections to extract multi-scale semantic features, while a transformer-based bottleneck module captures global contextual dependencies, including bilateral symmetry between left and right SNpc regions, to mitigate noise in biomarker extraction. We further establish a multi-task learning framework whose loss function uses two-dimensional Gaussian kernel-weighted masks to suppress discontinuous segmentation artifacts caused by repeated downsampling. Experiments on two datasets comprising 188 subjects from the Parkinson's Progression Markers Initiative and 40 subjects from London Health Sciences Center show that SNIB-TransNet outperforms six baseline methods, attaining a Dice similarity coefficient of 0.8691 and an AUC of 0.9439. Three-dimensional analysis across progression stages reveals only 12.3 percent degradation from healthy to severe PD, versus 18.7 percent for the strongest baseline, confirming robust clinical applicability across heterogeneous patient populations.