Revised NMS-driven pipeline for heart valve regurgitation and Kawasaki disease coronary aneurysm localization

Object detection in echocardiography is still uncommon, yet precise localization of pediatric valvular regurgitation and Kawasaki-related coronary aneurysms is critical. We introduce two lightweight variants of non-maximum suppression, rNMS-P (used only at inference) and rNMS-TP (used during both training and inference), that improve YOLO v5 to v9 detectors without altering their backbones. The system combines horizontal boxes for valve jets with rotation-aware oriented boxes for coronary segments, applies anatomical constraints by keeping a single high-confidence box per class, and can relax this rule for well-represented lesions. On color and grayscale echocardiograms, rNMS-TP increased mAP@.5 from 79.7% to 80.5% for regurgitation and from 85.7% to 86.5% for Kawasaki disease, with gains up to 2.3% at the stricter mAP@.5:.95 threshold; rNMS-P provided up to a 2.1% boost, all with negligible computational cost, offering a practical path toward explainable, operator-independent cardiac image assessment.