Purpose <p>Cardiovascular disease remains the leading global cause of annually rising morbidity and mortality, with myocardial fibrosis implicated in nearly all cardiac pathologies. Because early fibrosis often presents as a subclinical condition, accurate diagnosis is essential for timely intervention and improved prognosis. We therefore developed and validated a hybrid radiomics-deep learning framework that fuses multi-scale echocardiographic information to accurately assess myocardial fibrosis severity.</p> Methods <p>Myocardial fibrosis was induced in rabbits using the two-kidney, one-clip method. We collected 2,498 echocardiographic images from control, mild-fibrosis, and severe-fibrosis cohorts, delineating regions of interest (ROI). A novel architecture combined a U-Net backbone with a Multi-Scale Channel Attention Module to capture local and global textures. Nine feature-selection methods and nine machine-learning classifiers were exhaustively compared using 10-fold cross-validation; the optimal pipeline was identified using performance metrics.</p> Results <p>The fusion model outperformed standalone radiomics and deep-learning models. The configuration-GINI-based feature selection, RF classification, and MS-CAM modulus།exhibited the best discriminative ability, with an area under the curve (AUC) of 0.979.</p> Conclusions <p>Our work provides a pioneering and accurate approach for assessing the severity of myocardial fibrosis.</p>