MESTrans: Multi-scale embedding spatial transformer for medical image segmentation.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Medical Imaging

Background:

• Transformers utilizing self-attention mechanisms excel in computer vision tasks.
• Convolutional Neural Networks (CNNs) have limitations in capturing global and local information.
• Medical image segmentation is crucial for diagnosis and treatment planning.

Purpose of the Study:

• To propose a novel transformer-based medical image segmentation network, MESTrans.
• To enhance medical image segmentation by integrating self-attention mechanisms.
• To improve the adaptive extraction of discriminative information in medical images.

Main Methods:

• Developed the Multi-scale Embedding Spatial Transformer (MESTrans) network based on U-Net architecture.
• Introduced a Multi-scale Embedding Block (MEB) and Spatial Attention Transformer (SATrans) for dynamic receptive field adjustment.
• Implemented a Feature Fusion Module (FFM) for adaptive fusion of shallow and deep features.

Main Results:

• Achieved 81.23% Dice Similarity Coefficient (DSC) on the COVID-DS36 dataset.
• Obtained 89.95% DSC and 82.39% Intersection over Union (IoU) on the GlaS dataset.
• Demonstrated strong performance across CT, MR, and H&E-stained images, outperforming state-of-the-art methods.

Conclusions:

• MESTrans exhibits excellent generalization ability in medical image segmentation.
• The proposed model outperforms existing state-of-the-art methods.
• MESTrans is a promising tool for auxiliary clinical diagnosis and advancing medical intelligence.