(P276) ENHANCING CARDIOMYOPATHY DIAGNOSIS: A DEEP LEARNING APPROACH USING OXYGENATION-SENSITIVE CARDIAC MRI TO VISUALIZE MYOCARDIAL SCAR

Background: Cardiomyopathies, both ischemic and non-ischemic, pose a significant healthcare challenge due to their complex pathologies and the high prevalence of associated complications. Traditional diagnostic methods, such as cardiac MRI (CMR) with late gadolinium enhancement (LGE), are effective in detecting myocardial scars but require the use of contrast agents that can be harmful to certain patients. Oxygenation-sensitive CMR (OS-CMR) presents an innovative approach by eliminating the need for contrast altogether. This study aims to harness cutting-edge artificial intelligence technologies to develop a deep-learning framework for analyzing OS-CMR images. By leveraging transfer learning techniques, the project seeks to accurately visualize and classify pathological patterns inherent in both ischemic and non-ischemic cardiomyopathies, thus eliminating the need for contrast agents and reducing potential risks to patients.

METHODS AND RESULTS: Utilizing a dataset of 117 OS-CMR images annotated by expert cardiologists—24 non-ischemic, 25 ischemic, and 68 healthy—we developed a convolutional neural network (CNN) model. This model employs the Visual Geometry Group (VGG19) architecture and utilizes transfer learning to recognize and categorize patterns of myocardial scars. We optimized the model with an Adam optimizer and implemented a 0.5 dropout rate to prevent overfitting. For model validation and to enhance interpretability, we employed feature map visualization techniques. These techniques helped identify critical features extracted during learning, and heatmaps of these features were meticulously compared to scar locations identified on LGE images and matched to corresponding OS-CMR images, ensuring precise alignment. This comparative analysis highlights the model’s capability to accurately localize and differentiate between ischemic, non-ischemic, and healthy myocardial tissues, as shown in Figure 1 A and C. Throughout 48 epochs, the CNN model achieved a peak accuracy of 89.34%, with a precision of 90.52% and a recall of 86.07%. These metrics demonstrate the model's balanced effectiveness in predicting accurate classifications and ensuring reliability. The discriminative power of the model is further evidenced by a multi-class ROC AUC analysis, with individual class AUCs reported as 0.94 for class 0 (Blue Line), 0.92 for class 1 (Red Line), and 0.94 for class 2 (Green Line), as depicted in Figure 1 B.

Conclusion: Our project demonstrates the potential of deep learning to enhance the classification and precise localization of myocardial scars within Oxygenation-Sensitive CMR (OS-CMR) images. Future developments are necessary to validate this approach for routine clinical use in cardiovascular medicine.