Leveraging a Vision Transformer Model to Improve Diagnostic Accuracy of Cardiac Amyloidosis With Cardiac Magnetic Resonance

Abstract

Background

Cardiac magnetic resonance (CMR) imaging is an important diagnostic tool for diagnosis of cardiac amyloidosis (CA). However, discrimination of CA from other etiologies of myocardial disease can be challenging.

Objectives

The aim of this study was to develop and rigorously validate a deep learning (DL) algorithm to aid in the discrimination of CA using cine and late gadolinium enhancement CMR imaging.

Methods

A DL model using a retrospective cohort of 807 patients who were referred for CMR for suspicion of infiltrative disease or hypertrophic cardiomyopathy (HCM) was developed. Confirmed definitive diagnosis was as follows: 252 patients with CA, 290 patients with HCM, and 265 with neither CA or HCM (other). This cohort was split 70/30 into training and test sets. A vision transformer (ViT) model was trained primarily to identify CA. The model was validated in an external cohort of 157 patients also referred for CMR for suspicion of infiltrative disease or HCM (51 CA, 49 HCM, and 57 other).

Results

The ViT model achieved a diagnostic accuracy (84.1%) and an area under the curve of 0.954 in the internal testing data set. The ViT model further demonstrated an accuracy of 82.8% and an area under the curve of 0.957 in the external testing set. The ViT model achieved an accuracy of 90% (n = 55 of 61), among studies with clinical reports with moderate/high confidence diagnosis of CA, and 61.1% (n = 22 of 36) among studies with reported uncertain, missing, or incorrect diagnosis of CA in the internal cohort. DL accuracy of this cohort increased to 79.1% when studies with poor image quality, dual pathologies, or ambiguity of clinically significant CA diagnosis were removed.

Conclusions

A ViT model using only cine and late gadolinium enhancement CMR images can achieve high accuracy in differentiating CA from other underlying etiologies of suspected cardiomyopathy, especially in cases when reported human diagnostic confidence was uncertain in both a large single state health system and in an external CA cohort.