Mengyu Li, Magnús Magnússon, Thilo van Eimeren and Lotta M. Ellingsen
Introduction: Accurate brain MRI segmentation is essential for clinical and research applications, enabling detailed brain structure analysis and aiding neurological diagnoses. Manual segmentation remains the gold standard for accuracy but is extremely time-consuming. Recently, deep convolutional neural networks (CNNs) have emerged, offering high accuracy and rapid processing with smaller training datasets. A significant challenge in deep learning for medical imaging is managing the large data volumes relative to GPU memory. We present a multi-region-based 3D CNN brain segmentation algorithm designed for 12 deep-brain structures associated with Parkinson-plus syndromes imaging biomarkers. Despite limited GPU capacity, our method delivers highly accurate segmentation results with reduced training time.
Methods: Our method is implemented as a region-based approach. We divide the 12 targeted brain structures (four ventricular compartments, four brainstem sub-structures, putamen and caudate) into three focal regions according to their anatomical positions and relationships. This approach reduces training and processing times and enhances segmentation accuracy by optimizing each CNN for similar anatomical structures. The CNN was trained on 83 T1-weighted MRIs from three databases and tested on 40 semi-manually labeled MRIs.
Results: Our method provides robust segmentation results across multiple MRI scanners, achieving an average Dice Similarity Coefficient of 0.901 and a 95% Hausdorff Distance of 1.155 mm, indicating high segmentation accuracy.
Conclusions: We present a fully automated and robust brain segmentation method that shortens segmentation time from hours to seconds. Our method aims to facilitate the identification of novel imaging biomarkers to aid with early diagnosis of PPS and other neurodegenerative diseases.
