Hippocampal subfields segmentation in brain MR images using generative adversarial networks.

Area of Science:

• Medical Image Analysis
• Neuroimaging
• Artificial Intelligence

Background:

• Accurate segmentation of hippocampal subfields in brain MR images is crucial but challenging due to their small size and complex morphology.
• Traditional methods often struggle to achieve optimal results for these intricate structures.

Purpose of the Study:

• To develop and evaluate a novel method for accurate hippocampal subfield segmentation using generative adversarial networks (GANs).
• To enhance the precision of pixel-level classification for various hippocampal subfields in brain MR images.

Main Methods:

• Proposed a segmentation method employing a combination of a UG-net model for local feature extraction and an adversarial model for spatial consistency.
• Utilized alternate training of the UG-net and adversarial models to refine segmentation boundaries and ensure smooth class labels.
• Applied the method to segment CA1, CA2, DG, CA3, Head, Tail, SUB, ERC, and PHG subfields.

Main Results:

• Achieved high Dice Similarity Coefficients (DSC) for most subfields, with values up to 0.929.
• Demonstrated significant improvements over state-of-the-art methods, particularly for smaller subfields like ERC (20.93% increase) and PHG (16.30% increase).
• Showcased notable DSC increases for larger subfields, including Head (3.9%) and CA1 (9.03%).

Conclusions:

• The proposed GAN-based method significantly outperforms existing approaches for hippocampal subfield segmentation.
• The method provides superior results in segmenting both large and small hippocampal subfields, offering a valuable tool for neuroimaging research.