Improving Subcortical and Hippocampal Subfield Segmentation with a 3D Hybrid Deep Learning Solution
Aaron Cao, Zongyu Li, Jordan Jomsky, Andrew F. Laine, Jia Guo · Proceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition · 2025
Motivation: The most widely accepted brain segmentation tools like FreeSurfer are slow and inconvenient for large datasets. Faster deep-learning-based methods often sacrifice accuracy and reliability. Goal(s): We propose a novel deep-learning architecture for subcortical and hippocampal subfield segmentation and achieve consistent state-of-the-art performance. Approach: Our approach combines a 3D patch-based pipeline with a hybrid CNN-Mamba architecture, named MedSegMamba. Results: We evaluated MedSegMamba on FreeSurfer ground truths across various T1w MRI datasets. For subcortical segmentation, MedSegMamba consistently demonstrates strong performance over leading deep-learning alternatives, including CNN-Mamba, CNN-Transformer, and pure CNN networks. For hippocampal subfield segmentation, only MedSegMamba learned to segment all regions. Impact: Our proposed novel deep learning model, MedSegMamba, reliably demonstrated state-of-the-art segmentation performance and utility across numerous datasets. It outperformed other well-established deep learning tools on the difficult tasks of subcortical and hippocampal subfield segmentation. Code is available here: https://github.com/aaroncao06/MedSegMamba.