TPAFNet: Transformer-Driven Pyramid Attention Fusion Network for 3D Medical Image Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• 3D medical image segmentation is crucial for diagnosis.
• Current hybrid CNN-transformer networks have limitations in feature fusion.
• Effective fusion of channel and spatial attention is lacking.

Purpose of the Study:

• To introduce a novel multi-scale 3D medical image segmentation network, TPAFNet.
• To address limitations in existing hybrid network architectures.
• To improve the accuracy and efficiency of 3D medical image segmentation.

Main Methods:

• Developed TPAFNet, a hybrid CNN-transformer network.
• Utilized atrous convolution for multi-scale feature extraction.
• Introduced TPAF blocks for fused channel and spatial attention.
• Implemented TPAF connections in the decoder and a low-level encoding shortcut.
• Applied deep supervision with a CNN-based voxel-wise classifier.

Main Results:

• TPAFNet significantly outperformed state-of-the-art networks on two public datasets.
• The proposed TPAF blocks effectively fused multi-scale channel and spatial attention.
• The TPAF connection and shortcut preserved crucial low-level and original image features.
• Deep supervision enhanced network convergence.

Conclusions:

• TPAFNet offers a robust solution for 3D medical image segmentation.
• The novel attention fusion mechanism improves segmentation accuracy.
• This advancement aids clinicians in achieving more precise diagnoses.