Benchmarking 3D U-Net variants for volumetric segmentation of the substantia nigra and brain ageing related analysis using 7T MRI.

Abstract

Objective: Accurate segmentation of the substantia nigra (SN) is essential for assessing its structural integrity and volumetric changes in Parkinson's disease and brain aging. However, it remains challenging due to SN's small size, deep location, and low contrast on conventional MRI. This study benchmarked six U-Net variants baseline 3D U-Net, ResU-Net, SE-U-Net, Attention U-Net, U-Net++, and U-Net 3+ for SN segmentation using ultra-high-field 7 T MRI.

Methods: T1-weighted MP2RAGE images from the ATAG dataset (53 healthy subjects: 31 train, 7 validation and 15 test). Six architectures were evaluated: 3D U-Net, ResU-Net, SE-U-Net, Attention U-Net, U-Net++, and U-Net 3+.Performance was evalauted with Dice score, IoU, 95% HD95, ASSD, and also age-stratified analysis was considered for analysing the robustness of model across young, middle-aged, and elderly cohorts.

Results: While ResU-Net achieved the highest overlap metrics (Dice = 0.763, IoU = 0.618), where the baseline 3D U-Net outperformed all variants in boundary accuracy (HD95 = 13.85 mm, ASSD = 20.15 mm). The U-Net 3+ and SE-U-Net complex models relatively underperformed, indicated that architectural complexity does not enhance the segmentation performance of small subcortical structures. Descriptive Age-stratified analysis shows, ResU-Net maintained consistent performance across different age groups and 3D U-Net revealed strong age-invariance, however no statistically significant age-group differences were exhibited.

Conclusions: This study findings demonstrates architecture selection influences overlap, boundary detection and volume estimation small-structure segmentation and highlights the importance of evaluating model robustness to demographic variability. The findings offer a methodological benchmark for developing reliable SN volumetry tools for neurodegenerative research.