UMRFormer-net: a three-dimensional U-shaped pancreas segmentation method based on a double-layer bridged transformer...

Area of Science:

• Medical image analysis
• Deep learning for medical imaging
• Computational anatomy

Background:

• Transformer and Convolutional Neural Network (CNN) combinations show promise in medical image segmentation.
• Existing methods often use transformers as auxiliary modules, with limited investigation into optimal self-attention and convolution integration.

Purpose of the Study:

• To develop a novel deep learning framework for accurate 3D pancreas segmentation.
• To investigate the optimal integration of self-attention mechanisms with convolutional layers for enhanced medical image segmentation.

Main Methods:

• Designed a novel transformer block (MRFormer) integrating multi-head self-attention and residual depthwise convolution.
• Embedded the MRFormer block into the U-Net architecture (UMRFormer-Net) for 3D pancreas segmentation.
• Evaluated the framework on the CPTAC-PDA and Medical Segmentation Decathlon datasets.

Main Results:

• UMRFormer-Net achieved segmentation accuracy comparable or superior to state-of-the-art 3D methods.
• Demonstrated statistically significant improvements over existing transformer-related and 3D methods (P<0.05, P<0.01, or P<0.001).
• Achieved a higher Dice coefficient (85.54% and 77.36%) and lower 95% Hausdorff distance (4.05 and 8.34 mm).

Conclusions:

• UMRFormer-Net enhances pancreas segmentation accuracy by providing more precise boundary and region information.
• The novel MRFormer block effectively captures both long-range and local spatial information.
• The developed framework offers a significant advancement in 3D medical image segmentation, particularly for pancreatic structures.