T-Net: Nested encoder-decoder architecture for the main vessel segmentation in coronary angiography.

Area of Science:

• Medical Image Analysis
• Deep Learning Architectures
• Computer Vision

Background:

• U-Net architecture has limitations in feature propagation due to its single concatenate layer.
• Accurate segmentation of coronary angiography is crucial for cardiovascular disease diagnosis.

Purpose of the Study:

• To introduce T-Net, a nested encoder-decoder architecture designed to overcome U-Net's limitations.
• To evaluate T-Net's performance in segmenting main vessels in coronary angiography images.

Main Methods:

• Proposed T-Net, a convolutional neural network with multiple nested encoder-decoder blocks.
• Implemented T-Net and U-Net for comparative analysis on coronary angiography datasets.
• Optimized T-Net specifically for main vessel segmentation.

Main Results:

• T-Net achieved a Dice Similarity Coefficient (DSC) of 83.77%, outperforming U-Net by 10.69%.
• An optimized T-Net variant reached a DSC of 88.97%, surpassing U-Net by 15.89%.
• Activation visualizations confirmed T-Net's earlier mask prediction from lower decoder layers.

Conclusions:

• T-Net offers superior performance in medical image segmentation compared to U-Net.
• The nested architecture facilitates effective feature propagation for enhanced accuracy.
• T-Net shows potential for broader application in various medical image segmentation tasks.