MSA-MaxNet: Multi-Scale Attention Enhanced Multi-Axis Vision Transformer Network for Medical Image Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) excel at local feature extraction but struggle with long-range dependencies in medical image segmentation.
• Vision Transformers (ViTs) capture long-range dependencies but have limitations in leveraging local features.
• Existing hybrid models still face challenges with receptive fields and precise feature representation.

Purpose of the Study:

• To propose MSA-MaxNet, a novel encoder-decoder network for improved medical image segmentation.
• To enhance feature extraction by combining multi-axis self-attention (Max-SA) with CNN-based attention mechanisms.
• To address limitations in current segmentation models for complex medical imaging tasks.

Main Methods:

• Developed an encoder-decoder architecture utilizing MaxViT blocks with multi-axis self-attention (Max-SA) for comprehensive feature extraction.
• Incorporated a symmetric decoder with MaxViT blocks and upsampling layers to restore spatial resolution.
• Introduced a multi-scale convolutional block attention module (MCBAM) to refine skip connections and improve feature representation.

Main Results:

• Achieved state-of-the-art (SOTA) Dice scores of 85.59% on the Synapse dataset (multi-organ segmentation) and 95.26% on the Kvasir-SEG dataset (gastrointestinal polyp detection).
• Demonstrated strong performance on the ACDC dataset for cardiac analysis.
• Offered smaller model variants to cater to diverse computational constraints.

Conclusions:

• MSA-MaxNet provides a robust and effective framework for diverse medical imaging segmentation tasks.
• The model shows significant potential for applications in early cancer detection, cardiovascular disease diagnosis, and organ-level assessments.
• The integration of Max-SA and MCBAM modules offers a promising direction for advancing medical image analysis.