Pixel-Correlation-Based Scar Screening in Hypertrophic Myocardium.

The screening of myocardial scars in patients with hypertrophic cardiomyopathy(HCM) using cine magnetic resonance images(Cine-MRI) is critical in clinical practice. To tackle this task, we propose a locally-attentive pixel-correlation learning model. In such model, we first divide the image into different patches. Then, we employ a multi-layer perceptron to extract the scar features by modelling the correlation between pixels within each patch, as well as the correlation between pixels of different patches. To further enhance representation capability of the learned features, we incorporates an intra-attention mechanism to focus the pixels within the patches, and an inter-attention mechanism to focus the different the patches. After that, these learned deep features are combined with radiomics features and integrated into a linear classifier. The validation on clinical data shows powerful performance of our model, with accuracy, sensitivity, specificity, positive predictive value(PPV), negative predictive value(NPV), f1-score and area under the curve (AUC) values of 82.9%, 98.4%, 50.0%, 82.7%, 85.7%, 89.9% and 0.867, respectively. Our method has yielded promising experimental results, indicating that it can serve as a valuable tool for clinical physicians to screen scars.