PDAFormer 3+: A full-scale connected modified transformer with parallel dual attention for 3D medical image...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image segmentation is crucial for diagnostics, therapy, and research.
• Transformers offer advantages in capturing long-range dependencies, overcoming limitations of Convolutional Neural Networks (CNNs) in global context modeling.

Purpose of the Study:

• To propose PDAFormer 3+, a 3D medical image segmentation framework using a modified transformer with parallel dual attention.
• To enhance segmentation accuracy and computational efficiency for large-scale 3D medical images.

Main Methods:

• Introduced a parallel dual attention (PDA) mechanism for spatial and channel dependency modeling.
• Replaced transformer's MLP with a residual convolution block (RCB) for reduced complexity and enhanced local representations.
• Incorporated full-scale features and a convolution excitation module (CEM), inspired by U-Net 3+, with deep supervision.

Main Results:

• PDAFormer 3+ achieved high mean Dice Similarity Coefficients (DSC): 86.90% on Synapse, 92.54% on ACDC, and 91.93% on type-B AD.
• Demonstrated significant reduction in computational burden for large-scale 3D medical images.
• Maintained strong efficiency alongside high accuracy.

Conclusions:

• PDAFormer 3+ effectively integrates CNNs and transformers, leveraging local details and global context for accurate and efficient 3D medical image segmentation.
• The framework shows promise for improving diagnostic accuracy and healthcare efficiency.