HGM-cNet: Integrating hippocampal gray matter probability map into a cascaded deep learning framework improves...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate hippocampus segmentation is vital for diagnosing neuropsychiatric disorders like Alzheimer's disease (AD).
• Existing segmentation methods face challenges in precision and generalizability.

Purpose of the Study:

• To develop a robust cascaded deep learning framework, HGM-cNet, for improved hippocampus segmentation.
• To integrate a hippocampal gray matter (HGM) probability map into the deep learning architecture.

Main Methods:

• Developed HGM-cNet, a cascaded deep learning framework using two identical convolutional neural networks (CNNs).
• Each CNN incorporated Attention Block, Residual Block, and DropBlock within an encoder-decoder architecture with skip-connections.
• Integrated HGM probability map with feature maps from the first CNN.

Main Results:

• HGM-cNet outperformed seven multi-atlas and six deep learning methods on 135 T1-weighted MRI scans.
• Achieved average Dice scores >0.89 for both left and right hippocampus, exceeding other methods by ~1%.
• Demonstrated superior performance across cognitive normal, mild cognitive impairment, and AD groups, and validated generalizability with various CNN models.

Conclusions:

• The cascaded deep learning framework with an integrated HGM probability map significantly improves hippocampus segmentation.
• HGM-cNet facilitates more accurate capture of hippocampal atrophy in AD analysis.
• The framework demonstrates stability, convenience, and generalizability for hippocampus segmentation.