HMDA: A Hybrid Model with Multi-scale Deformable Attention for Medical Image Segmentation

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) struggle with global feature extraction in medical imaging.
• Transformers offer long-range modeling but can be computationally redundant with high-dimensional data.
• Integrating CNNs' local detail with Transformers' global context is challenging.

Purpose of the Study:

• To develop an improved medical image segmentation model addressing limitations of existing CNN and Transformer architectures.
• To enhance the model's ability to focus on salient image regions and integrate multi-scale features effectively.

Main Methods:

• Proposed a Hybrid Transformer and CNN architecture (HMDA).
• Introduced a Multi-scale Spatially Adaptive Deformable Attention (MSADA) mechanism for focused attention on key sampling points.
• Developed a Cross Attention Bridge (CAB) module for integrating multi-scale Transformer and local CNN features via channelwise cross-attention.

Main Results:

• HMDA demonstrated effectiveness across multiple medical image segmentation datasets.
• The proposed MSADA mechanism improved attention efficiency and focus on salient regions.
• The CAB module successfully enriched feature synthesis by integrating diverse feature types.

Conclusions:

• The HMDA architecture offers a promising approach for medical image segmentation, outperforming previous methods.
• The novel attention and feature integration mechanisms contribute to enhanced segmentation performance.
• This hybrid approach effectively balances local and global feature extraction for medical imaging tasks.