Medical image segmentation using boundary-enhanced guided packet rotation dual attention decoder network.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Automatic medical image segmentation is crucial but challenged by complex backgrounds, unclear boundaries, and variable organ sizes, leading to feature loss and low accuracy.
• Existing methods struggle to preserve deep feature information and accurately delineate organ boundaries in medical images.

Purpose of the Study:

• To develop a deep learning network that effectively preserves image deep feature information.
• To improve medical image segmentation accuracy, particularly for cases with unclear boundaries.

Main Methods:

• Developed BGRANet, a deep learning framework incorporating a packet rotation convolutional fusion encoder for feature extraction.
• Introduced a boundary-enhanced guided packet rotation dual attention decoder to refine segmentation maps and integrate prior information.
• Utilized a multi-resolution fusion module to generate high-resolution feature maps for enhanced segmentation.

Main Results:

• BGRANet demonstrated superior segmentation performance on the CHAOS (91.73% Dice for 4 classes) and Herlev (Dice 98.08% for 2 classes) datasets.
• The model achieved high prediction, sensitivity, specificity, accuracy, and Dice scores across different classification tasks.
• Experimental results confirm BGRANet's effectiveness in improving medical image segmentation outcomes.

Conclusions:

• The proposed boundary-enhanced guided packet rotation dual attention decoder network significantly improves segmentation accuracy.
• BGRANet achieves high segmentation accuracy while utilizing a reduced number of parameters, indicating an efficient model design.