UNet based on dynamic convolution decomposition and triplet attention.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Medical image segmentation faces challenges in robustness and generalization across diverse diseases and imaging modalities.
• While U-Net is popular for medical image segmentation, limitations in feature expression and segmentation accuracy persist.
• Existing methods struggle with effectively highlighting subtle lesion regions amidst complex image data.

Purpose of the Study:

• To develop an improved deep learning model for medical image segmentation that addresses the limitations of existing U-Net architectures.
• To enhance feature extraction capabilities and improve the accuracy of lesion segmentation.
• To create a versatile model applicable to various medical imaging tasks and datasets.

Main Methods:

• Proposed DTA-UNet, integrating Dynamic Convolution Decomposition (DCD) and Triple Attention (TA) mechanisms into an Attention U-Net baseline.
• DCD replaces conventional convolutions to boost feature extraction efficiency.
• TA combined with Attention Gates (AG) refines skip connections, reducing redundant information for precise lesion highlighting.

Main Results:

• DTA-UNet demonstrated significant improvements in segmentation metrics across COVID-SemiSeg, ISIC 2018, and a clinical stroke dataset.
• Ablation studies confirmed the effectiveness of DCD and TA, with only a marginal increase in parameters (0.7628 M) compared to the baseline.
• The model exhibited universality by performing well on diverse image types, outperforming eight state-of-the-art methods.

Conclusions:

• DTA-UNet effectively enhances feature extraction and lesion segmentation accuracy in medical imaging.
• The proposed DCD and TA modules offer a computationally efficient way to improve segmentation performance.
• DTA-UNet presents a robust and universally applicable solution for various medical image segmentation challenges.