1357789 - Deep Learning using Cardiovascular Magnetic Resonance Imaging Provides Incremental Benefit in the Diagnosis of Cardiac Amyloidosis

Deep learning (DL) using cardiovascular magnetic resonance imaging (CMR) is becoming increasing prevalent. However, few studies have attempted to characterize the potential benefit of DL methods for disease classification compared to physician diagnostic interpretation. In this study, we seek to demonstrate the accuracy of deep learning classification compared to physician interpretation of CMR images in diagnosing cardiac amyloidosis (CA) in a retrospective analysis.

Methods:

A retrospective cohort of 104 patients with a definitive diagnosis of CA, determined by endomyocardial/fat pad biopsy, serum light chains, and/or PYP scan, who underwent CMR cine imaging at single institution between 2005-2021 were selected for analysis. For each patient, lexicon used in the imaging report completed at the time of the MRI was used to characterize the accuracy and confidence of the physician’s image-based diagnosis. Patients were stratified into five categories – high confidence, moderate confidence, and low confidence of cardiac amyloidosis, no diagnosis given, or an incorrect diagnosis. The relative accuracy of a DL DenseNet121 convolutional neural network trained to distinguish CA and hypertrophic cardiomyopathy phenotypes using CMR cine images on a separate training dataset was assessed in each of the confidence levels.

Results:

The physician CMR interpretation accurately diagnosed CA in 84/104 patients (80.8%), had no diagnosis in 4/104 patients (3.7%), and had an incorrect diagnosis in 16/104 patients (15.1%). The most common reason for no diagnosis or an incorrect diagnosis was poor quality late gadolinium enhancement images. The DL network was able to correctly predict CA using a single mid ventricular short axis slice CMR cine images in 78/84 patients (92.9%) with a low, moderate, or high confidence diagnosis based on the CMR imaging report. The DL network was able to correctly predict CA in 16/20 (80%) of patients with no diagnosis or an incorrect diagnosis.

Conclusion:

The DL network using CMR cine images exhibited high accuracy across all confidence levels when compared to physician image-based diagnosis. This demonstrates the potential incremental benefit of DL networks as a clinical decision support tool in CA.