(P273) DEEP LEARNING-BASED ANALYSIS OF BSSFP CMR IMAGES FOR THE DETECTION OF HYPERTROPHIC CARDIOMYOPATHY

Background: Hypertrophic Cardiomyopathy (HCM) is a prevalent genetic cardiovascular disorder, distinguished by thickened heart muscles and distinct fibrosis patterns within the myocardium that could lead to heart failure. Although Late Gadolinium Enhancement (LGE) in cardiovascular magnetic resonance (CMR) remains the gold standard for fibrosis detection in HCM, it poses certain limitations, including post-injection retention in the body. Consequently, there is a growing interest in exploiting needle-free CMR sequences such as short-axis (SAX) cine MRI images for precise diagnosis and evaluation of HCM. This study leverages AI to enhance diagnostic accuracy in HCM by identifying pathological patterns in the myocardium. Advanced deep learning techniques distinguish between fibrosis and healthy tissue, analyzing abnormalities akin to those in LGE imaging. We also visualize feature maps to display fibrosis patterns learned by the AI, refining the precision of non-invasive cardiology diagnostics.

METHODS AND RESULTS: In this study, we employed a transfer learning approach using a 3D ResNet architecture, tailored to analyze both systolic and diastolic phases of HCM from 10-slice SAX cine MRI images of 23 HCM patients and 28 healthy individuals. Leveraging the deep residual learning framework of the ResNet, pre-trained on the ImageNet database, we achieved robust feature extraction essential for distinguishing between pathological and normal myocardial tissues. The model was optimized with the Adam optimizer at a learning rate of 0.000295, incorporating a 50% dropout rate to enhance generalization and prevent overfitting. After training for 50 epochs, the model demonstrated exceptional discriminative capability, achieving a validation accuracy, precision, and recall of 92.71% each, complemented by an outstanding ROC-AUC score of 0.9929 (Figure 1). Significantly, we utilized heat maps generated from AI feature maps to visualize and identify fibrosis patterns in the cine images, which closely matched the abnormalities observed in Late Gadolinium Enhancement (LGE), the current gold standard (Figure 1. C). These results underscore the potential of advanced AI networks to enhance diagnostic accuracies for cardiac conditions.

Conclusion: Our study demonstrated that a 3D ResNet architecture can effectively differentiate between HCM and normal cases by analyzing cardiac cine MRI images during the systolic and diastolic phases. The AI model successfully identified pathological patterns that closely correspond with those observed in Late Gadolinium Enhancement (LGE), the current gold standard. This finding suggests the potential for AI-driven models trained on needle-free images to transform cardiac diagnostics in clinical settings by eliminating the need for contrast injections