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

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Early diagnosis of neurodegenerative diseases is challenging due to subtle, overlapping symptoms.
• Automated MRI segmentation is vital for detecting subtle brain changes, but manual methods are impractical.
• Deep learning methods face GPU memory constraints with large datasets.

Purpose of the Study:

• To develop an efficient deep learning MRI segmentation method for 12 deep-brain structures relevant to Parkinson Plus Syndromes.
• To optimize GPU usage and reduce training times for large-scale brain image analysis.
• To improve the accuracy and robustness of automated brain segmentation for clinical applications.

Main Methods:

• A region-based U-net deep learning architecture was employed.
• Brain images were divided into targeted regions (brainstem, ventricular system, striatum) to optimize processing.
• The method was validated on three datasets, including a 660-subject clinical cohort.

Main Results:

• Achieved superior segmentation performance with a mean Dice Similarity Coefficient (DSC) of 0.90.
• Demonstrated high accuracy with a 95% Hausdorff Distance (HD95) of 1.35 mm and Average Symmetric Surface Distance (ASSD) of 0.45 mm.
• Significantly reduced training time from days to hours and processing time to under a second per subject.

Conclusions:

• The proposed region-based U-net method offers accurate, robust, and efficient MRI segmentation for deep-brain structures.
• This approach overcomes GPU memory limitations and accelerates training for large datasets.
• The method shows practical applicability for differentiating diseases and aiding early diagnosis in clinical settings.