EPT-Net: Edge Perception Transformer for 3D Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) excel in medical image segmentation but struggle with long-range dependencies.
• Transformers offer global prediction but may lack low-level feature detail for precise segmentation.
• High-resolution 3D medical image processing demands significant computational resources.

Purpose of the Study:

• To develop an accurate medical image segmentation method that overcomes limitations of existing CNN and Transformer models.
• To enhance the modeling of long-range dependencies and the capture of fine-grained edge information.
• To propose an efficient network architecture for high-resolution 3D medical image segmentation.

Main Methods:

• Proposed EPT-Net, an encoder-decoder network integrating edge perception and Transformer architecture.
• Introduced a Dual Position Transformer to improve 3D spatial positioning capabilities.
• Developed an Edge Weight Guidance module to extract critical edge information from low-level features without increasing parameters.

Main Results:

• Demonstrated effectiveness on SegTHOR 2019, Multi-Atlas Labeling Beyond the Cranial Vault, and KiTS19-M datasets.
• Achieved significant improvements in medical image segmentation compared to state-of-the-art methods.
• Validated the network's ability to accurately segment medical images, particularly concerning edge details.

Conclusions:

• EPT-Net effectively combines Transformer and CNN strengths for superior medical image segmentation.
• The proposed Dual Position Transformer and Edge Weight Guidance modules enhance accuracy and efficiency.
• EPT-Net represents a significant advancement in accurate and detailed medical image segmentation.