Enhancing Cardiac Ultrasound Regurgitation Detection Using YOLOv9 and Integrated Attention Modules

In this study, we examine a robust framework for detecting regurgitation in cardiac ultrasound images based on the YOLOv9 architecture. Our approach leverages a dataset validated using a 5-fold cross-validation strategy to ensure reliable and generalizable performance. To enhance detection accuracy and feature representation, we integrate nine well-known attention modules-namely, Squeeze-and-Excitation (SE), Con- volutional Block Attention Module (CBAM), SimAM, A2-NET, SKAttention, Efficient Attention, Parnet Attention, NAMAttention, and Coordinate Attention-into the network architecture. These modules are strategically embedded both within the backbone and the detection head, with the latter configured to accept either one or three attention positions. During extensive experimentation, we observed that the Efficient Attention and A2-NET modules required careful modification to operate effectively. Importantly, among the various configurations tested, CBAM demonstrated superior performance, achieving the highest mAP@50, which indicates its exceptional capability in enhancing the detection of regurgitation regions in cardiac ultrasound images. This work underscores the potential benefits of integrating specialized attention mechanisms into modern deep learning frameworks for improved medical image analysis.