EMCAH-Net: an effective multi-scale context aggregation hybrid network for medical image segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image segmentation is challenging due to variable scales, complex shapes, and low contrast.
• Existing hybrid networks struggle to integrate local (CNN) and global (Transformer) features effectively.
• Self-attention mechanisms in current models often neglect crucial spatial and channel information.

Purpose of the Study:

• To develop a robust hybrid deep learning model for accurate medical image segmentation.
• To enhance the integration of local multi-scale and global features in medical image analysis.
• To improve segmentation of challenging features like small targets and blurred boundaries.

Main Methods:

• Proposed the effective multi-scale context aggregation hybrid network (EMCAH-Net).
• Integrated an effective multi-scale context aggregation (EMCA) block for local feature encoding.
• Employed a dual-attention augmented self-attention (DASA) block to enhance global representation via spatial and channel attention.

Main Results:

• Achieved superior Dice similarity coefficient (DSC) scores on Synapse (84.73%), ACDC (92.33%), and DRIVE (82.47%) datasets.
• Demonstrated significant improvements over state-of-the-art methods, outperforming TransUNet by 7.25% DSC on the Synapse dataset.
• Maintained computational efficiency, requiring only 25% of parameters and 71% of FLOPs compared to TransUNet.

Conclusions:

• EMCAH-Net excels at segmenting multi-scale, small, and boundary-blurred features in medical images.
• The model surpasses existing CNN, Transformer, and hybrid architectures in abdominal, cardiac, and retinal image segmentation.
• EMCAH-Net offers a robust and efficient solution for medical image segmentation tasks.