Region-based U-nets for fast, accurate, and scalable deep brain segmentation: Application to Parkinson Plus Syndromes

The early diagnosis of age-related neurodegenerative diseases, which often progress to dementia, poses significant clinical challenges due to subtle and overlapping symptoms of these diseases at early stage. Automated MRI segmentation is important for early detection, as it offers consistent measurements and the ability to detect subtle structural changes in the brain. Manual segmentation is impractical for large datasets or clinical use. Deep learning approaches provide fast processing, however, they often encounter graphics processing unit (GPU) memory constraints when handling large datasets. Here we introduce a deep learning-based approach using region-based U-nets specifically designed to segment 12 deep-brain structures relevant to Parkinson Plus Syndromes. By dividing the brain image into targeted regions around the brainstem, ventricular system, and striatum, our method optimizes GPU usage and significantly reduces training times, while maintaining high accuracy. Validating the proposed method on three datasets, including a 660-subject clinical dataset comprising both healthy controls and patients with various movement disorders, we demonstrate robustness and practical applicability in separating different diseases. The method achieves superior segmentation performance compared to state-of-the-art methods, with a mean Dice Similarity Coefficient (DSC) of 0.90, a 95% Hausdorff Distance (HD95) of 1.35 mm, and an Average Symmetric Surface Distance (ASSD) of 0.45 mm, showcasing its segmentation accuracy and robustness. Furthermore, our method outperforms these methods by reducing training time from several days to a few hours while providing a processing time of less than a second per subject. The source code and trained model will be made publicly available on GitHub.