MSCMH-Net: A multi-scale channel-mixing hybrid network for whole-brain segmentation

Area of Science:

• Medical Image Analysis
• Neuroscience
• Artificial Intelligence

Background:

• Whole-brain segmentation is crucial for quantitative assessment of brain regions in clinical practice and research.
• Challenges include numerous brain regions, inter-class heterogeneity, and complex spatial dependencies.
• Accurate segmentation requires local feature delineation and modeling of long-range/global dependencies.

Purpose of the Study:

• To develop an advanced deep learning framework for accurate whole-brain segmentation.
• To address the limitations of existing methods in capturing both local and global contextual information.
• To propose the Multi-Scale Channel-Mixing Hybrid Network (MSCMH-Net).

Main Methods:

• Developed MSCMH-Net, a hybrid Convolutional Neural Network (CNN)-Multilayer Perceptron (MLP) framework.
• Integrated CNNs for local feature extraction and MLPs for long-range dependency modeling.
• Employed a channel-mixing module with exponential moving average (EMA) fusion for integrating global and local information.

Main Results:

• MSCMH-Net demonstrated competitive performance on a composite dataset of 106 brain MR scans.
• The hybrid approach effectively balances local feature capture and global context modeling.
• Validated through comprehensive experiments on datasets including MICCAI-2012, ADNI, and OASIS.

Conclusions:

• MSCMH-Net offers a promising approach for accurate whole-brain segmentation.
• The hybrid CNN-MLP architecture effectively models complex brain structures and dependencies.
• The findings support the utility of MSCMH-Net in medical imaging and neuroscience research.