Digital Poster 11: Prediction of Ischemic Cardiomyopathy Using a Deep Neural Network with Non-contrast Cine Cardiac Magnetic Resonance Images

Late gadolinium enhancement (LGE) Cardiovascular Magnetic Resonance (CMR) imaging is an established clinical method for assessing irreversible myocardial tissue damage. Although this method offers important diagnostic information, it requires the administration of a gadolinium-based contrast agent, adding time and complexity to the CMR exam. Given these concerns, there is a need to explore alternative approaches for assessing myocardial tissue damage. Past studies suggest that non-contrast cine CMR images have the potential to additionally characterize myocardial tissue. This potential can be enhanced using deep neural networks (DNNs), a class of artificial neural networks designed to model complex nonlinear relationships between inputs and outputs. This study aimed to train a DNN that can correctly classify non-contrast cine CMR images as belonging to either an ischemic cardiomyopathy (ICMP) patient cohort or a normal healthy cohort with a fully automated pipeline.

Methods:

We identified patients with ICMP and healthy volunteers having undergone a CMR exam. All scans were conducted using clinical 3T MRI systems. Each patient's cine images were acquired as a stack of short-axis slices and converted into a single 3D NumPy array. All patient arrays were preprocessed by normalization based on standard deviation and padded to a uniform size of 512 x 512 x 475 (Figure 1A). ResNet50, a 50-layer pre-trained deep learning model, was adapted and modified to receive the inputs with 512 x 512 x 475 pixels. The model was then trained on the non-contrast CMR cine image arrays to output an appropriate binary patient classification of “ICMP” or “Normal”.

Results:

The study included 81 ICMP patients (69 males and 12 females, mean age 61.9 ± 13.1 years) and 121 normal subjects (57 males and 64 females, mean age 44.4 ± 13.2 years). The trained model achieved an AUC score of 0.88 (Figure 1B), demonstrating effective discrimination between the Normal and ICMP classes. The trained model also attained a validation accuracy of 80.0% (Figure 1C) and a validation loss of 0.753 (Figure 1D). Throughout the training, both the training and validation accuracies were maximized over time, while both the training and validation losses were minimized. These outcomes indicate that the model was trained properly.

Conclusion:

The proposed fully automated DNN algorithm can correctly classify patients with ischemic cardiomyopathy from normal subjects in standard, native cine stacks. Further research is needed to assess the individual diagnostic accuracy in various clinical scenarios and to apply this approach to differentiate between cardiac diseases.