A study highlights the potential of a novel deep learning (DL) model that uses vision transformer (ViT) architecture to enhance the diagnostic accuracy of cardiac amyloidosis (CA) through cardiac magnetic resonance (CMR) imaging. Research published in JACC: Cardiovascular Imaging outlines the case for the use of artificial intelligence (AI) to improve CA differentiation from other cardiac conditions, such as hypertrophic cardiomyopathy (HCM). The US-based team added that AI could prove especially valuable where human diagnostic confidence is low, or where imaging resources are limited. “The DL algorithm was able to achieve an accuracy of 61.1% in patients in this internal test set sub cohort [with uncertain or incorrect diagnoses],” said the authors of the study. “Furthermore, removing cases with poor scan quality, isolated CA without clinical progression, and patients with dual diagnoses resulted in a substantial increase in accuracy...” “This result suggests that an AI clinical decision support tool may play a valuable assistive role in augmenting clinical diagnosis of CA, particularly when there is uncertainty or limited expert experience." Key findings Led by Joshua Cockrum, MD, from the University of Michigan Hospitals in Ann Arbor, the investigation noted a ViT model diagnostic accuracy of 84.1% and an area under the curve of 0.954 in the internal testing data set. In the external testing set, the model achieved an accuracy of 82.8% and an area under the curve of 0.957. The model went on to achieve an accuracy of 90% (n=55 of 61) among studies with clinical reports noting a 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. Further findings revealed the DL accuracy of this cohort increased to 79.1% when twelve studies with poor image quality, dual pathologies or ambiguity of clinically significant CA diagnosis were removed. “The ViT architecture is a DL architecture based on multihead attention made popular by natural language processing in which images are broken into patches and treated like sequences similar to sentences,” explained the authors. Future implications The investigation, which was also published online, touched on the broader implications of the study’s findings commenting on the transformative role of AI in cardiac imaging workflows. This includes the model’s ability to accurately distinguish CA from similar conditions, like hypertrophic cardiomyopathy (HCM), which could reduce diagnostic delays, expediting treatment initiation for patients with CA. Another benefit was the way AI tools could complement human expertise, offering recommendations or probabilities of diagnoses that could aid clinicians to make more informed decisions. “Given that the model was tuned to provide maximum accuracy across all 3 classes, it would be possible to increase sensitivity to CA, although at the cost of decreased positive predictive value,” said the paper’s authors. Despite the promise of such technology, the study highlighted a number of limitations, namely the variations in imaging protocols and patient populations that emphasized the importance of personalizing AI tools to certain clinical environments. While the model performed well in external datasets, the research team stressed the need for prospective studies to further support its efficiency in real-world clinical settings. Study approach A DL model using a retrospective cohort of 807 patients (male: 66%) who were referred for CMR for suspicion of infiltrative disease or hypertrophic cardiomyopathy (HCM) was developed. Definitive diagnosis confirmed 252 patients with CA, 290 patients with HCM and 265 with neither CA nor HCM (other). This cohort was split 70/30 into training and test sets (Ohio cohort). A ViT model was trained primarily to identify CA. The model was validated in an external cohort of 157 patients (Florida cohort) also referred for CMR for suspicion of infiltrative disease or HCM (51 CA, 49 HCM, 57 other). The ViT model’s performance was compared to traditional methods and human diagnostic accuracy. Source: Cockrum J, Nakashima M, Ammoury C, et al. Leveraging a Vision Transformer Model to Improve Diagnostic Accuracy of Cardiac Amyloidosis With Cardiac Magnetic Resonance. JACC Cardiovasc. Imaging. 2024 (Article in Press). Image Credit: HadK – stock.adobe.com