QueryCLR: a Query-based Contrastive Visual Representation Learning for Echocardiogram Classification

Echocardiography is a leading cardiac imaging technique that requires substantial expertise when performing an analysis, and manual labeling is challenging due to various acquisition devices. This research uses an innovative query-based contrast learning visual representation methodology called QueryCLR. By integrating CNN's precise feature extraction with transformer models' broad contextual understanding, QueryCLR substantially enhances the stability and accuracy of image analysis. Our experimental study, conducted on 12 different two-dimensional echocardiogram datasets, validates QueryCLR's effectiveness. The method achieved a maximum accuracy of 79.4%, showing a 1.2% superiority over existing self-supervised models. Moreover, by just using 10% of labeled data via transfer learning, QueryCLR achieved 84.4% accuracy, surpassing the supervised learning model by 0.9%. These results highlight the promising potential of self-supervised learning in enhancing accuracy and efficiency in cardiac disease identification through echocardiographic classification while providing new insights for future research.