SSMCE: A semi-supervised learning framework for myocardial segmentation in myocardial contrast echocardiography

Accurate myocardial segmentation in myocardial contrast echocardiography (MCE) images remains challenging due to the scarcity of publicly available labeled datasets and the pervasive presence of speckle noise. Currently, echocardiographers must manually delineate myocardial contours, a clinical workflow step that is both labor-intensive and prone to variability. To address these limitations, we propose SSMCE, a novel semi-supervised learning framework specifically designed for myocardial segmentation in MCE images. The proposed framework adopts a tri-model architecture comprising two structurally distinct student models and an adaptively assembled teacher model. This design inherently introduces model-level perturbations to promote output diversity, thereby reducing overfitting and improving generalization performance. In addition, a specialized loss function is designed to guide the model's self-correction behavior by increasing uncertainty in misclassified bias regions and reinforcing confidence in accurate ones, facilitating convergence. Experimental results on our selfconstructed dataset demonstrate that the proposed loss function improves the primary evaluation metric by 1.75%. Furthermore, the proposed method achieves state-of-the-art performance when compared with existing approaches. The results demonstrate that SSMCE provides a robust and efficient approach for rapid myocardial detection and precise segmentation, offering significant potential to streamline clinical workflows in MCE imaging.