Enhanced coronary artery segmentation and stenosis detection: Leveraging novel deep learning techniques

Abstract

Coronary artery disease (CAD) is a significant global health concern, emphasizing the need for reliable and automated diagnostic solutions. This study proposes a novel deep learning framework aimed at improving both full artery segmentation and stenosis localization by incorporating Self-Organizing Neural Networks (Self-ONN). The DenseSelfU-Net model leverages DenseNet121 as an encoder and a Self-ONN enhanced decoder within a U-Net based architecture to achieve robust feature extraction and precise full artery segmentation, achieving an IoU of 82.52% and a Dice score of 90.35% on the ARCADE Challenge dataset. For stenosis localization, Self-ONN is integrated into key components of the Multi-Scale Attention Network (MA-Net), which includes the Multi-Scale Fusion Attention Block (MFAB) and the Position-wise Attention Block (PAB), capturing complex vascular patterns through both local and global dependencies and resulting in the DenseSelfMA-Net model. The DenseSelfMA-Net achieves Dice scores of 60.59% and 60.36% and IoU scores of 46.09% and 45.36% for the MFAB and PAB configurations, respectively on the ARCADE challenge dataset. These results demonstrate the effectiveness of Self-ONN in enhancing diagnostic precision and facilitating early CAD diagnosis, with promising implications for clinical practice.