BATFormer: Towards Boundary-Aware Lightweight Transformer for Efficient Medical Image Segmentation

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) have limitations in receptive fields for medical image segmentation.
• Transformers offer potential for global representation learning but face challenges in computational complexity and rigid window partitioning.
• These challenges hinder the effective deployment of transformers in medical image segmentation tasks.

Purpose of the Study:

• To address the computational complexity and rigid window partitioning issues in transformers for improved medical image segmentation.
• To develop a customized transformer architecture that enhances both efficiency and performance in segmenting medical images.

Main Methods:

• Proposed a boundary-aware lightweight transformer (BATFormer) incorporating a cross-scale global transformer (CGT) module for efficient global feature extraction.
• Introduced a boundary-aware local transformer (BLT) module with adaptive window partitioning guided by entropy to preserve shape and reduce complexity.
• CGT module jointly utilizes multiple small-scale feature maps for richer global features with lower computational cost.

Main Results:

• BATFormer achieved superior performance on the ACDC dataset with Dice scores of 92.84% (average), 91.97% (right ventricle), 90.26% (myocardium), and 96.30% (left ventricle).
• On the ISIC 2018 dataset, BATFormer obtained the best Dice, IoU, and ACC scores of 90.76%, 84.64%, and 96.76%, respectively.
• BATFormer demonstrated the lowest model parameters and computational complexity compared to existing state-of-the-art methods.

Conclusions:

• The study highlights the need for customized transformers tailored for efficient and effective medical image segmentation.
• BATFormer's design offers a promising and extendable framework for various medical imaging applications.
• The adaptive windowing and cross-scale interaction strategies are key to achieving high performance with reduced computational load.