SWRAU-Net: Shearlet-integrated Convolution Based Wavelet Residual Attention U-Net for Segmentation of Carotid Artery Vessel Wall in 3D-MERGE MRI

Abstract Atherosclerosis is a critical disease that poses health risks such as stroke, heart attack, etc. Atherosclerosis is detected by measuring of the lumen area within the carotid vessels. 3D Motion-Sensitized Driven-Equilibrium prepared Rapid Gradient Echo (3D-MERGE) is a fast and high-resolution imaging technique that is promising for detection of atherosclerosis. However, very few research works explore 3D-MERGE images for carotid wall segmentation. In this paper, we propose SWRAU-Net model, designed for accurate segmentation of carotid vessel walls in 3D-MERGE images. The SWRAU-Net model is a modified U-Net model, which integrates Shearlet transform into the convolution block, and Haar wavelet transform into the U-Net architecture with a residual backbone and soft attention gate. Integrating the Shearlet transform helps extract high-contextual and directional vessel characteristics more effectively, as compared to wavelet transform. SWRAU-Net is trained with data of 25 patients obtained from the Carotid Vessel Wall segmentation challenge and validated on test data comprising of another 25 patients. The model was investigated with different loss functions and optimization algorithms. SWRAU-Net using Unified Focal loss function and Root Mean Square propagation shows the best results with a high Dice score, sensitivity, and specificity measures of approximately 77%, 83% and 99%, respectively. SWRAU-Net demonstrates better performance compared to existing state-of-the-art models based on U-Net. Thus, the proposed SWRAU-Net can serve as an effective segmentation framework for carotid vessel wall delineation and can facilitate clinicians to precisely assess disease progression and tailor treatment strategies accordingly.