January 1, 2020Open Access
Deep learning to diagnose cardiac amyloidosis from cardiovascular magnetic resonance
Structured PICO
Does a deep learning strategy evaluating LGE images accurately diagnose cardiac amyloidosis compared to a machine learning algorithm using manually extracted features in patients with suspected cardiac amyloidosis?
P
Population
206 subjects with suspected cardiac amyloidosis (n = 100, 49% with unexplained left ventricular hypertrophy; n = 106, 51% with blood dyscrasia and suspected light-chain amyloidosis)
I
Intervention
Deep learning (DL) strategy evaluating short axis, 2-chamber, 4-chamber late gadolinium enhancement (LGE) images from 1.5 T CMR
C
Comparator
Machine learning (ML) algorithm considering all manually extracted features (LV volumes, mass and function, LGE pattern, early blood-pool darkening, pericardial and pleural effusion, etc.)
O
Outcome
Diagnostic accuracy for cardiac amyloidosis (measured by AUC, precision, recall score, and F1 score)surrogate
A deep learning approach using LGE CMR images provides high diagnostic accuracy for cardiac amyloidosis, comparable to machine learning models requiring manual feature extraction.
Abstract
BACKGROUND: Cardiovascular magnetic resonance (CMR) is part of the diagnostic work-up for cardiac amyloidosis (CA). Deep learning (DL) is an application of artificial intelligence that may allow to automatically analyze CMR findings and establish the likelihood of CA. METHODS: 1.5 T CMR was performed in 206 subjects with suspected CA (n = 100, 49% with unexplained left ventricular (LV) hypertrophy; n = 106, 51% with blood dyscrasia and suspected light-chain amyloidosis). Patients were randomly assigned to the training (n = 134, 65%), validation (n = 30, 15%), and testing subgroups (n = 42, 20%). Short axis, 2-chamber, 4-chamber late gadolinium enhancement (LGE) images were evaluated by 3 networks (DL algorithms). The tags "amyloidosis present" or "absent" were attributed when the average probability of CA from the 3 networks was ≥ 50% or < 50%, respectively. The DL strategy was compared to a machine learning (ML) algorithm considering all manually extracted features (LV volumes, mass and function, LGE pattern, early blood-pool darkening, pericardial and pleural effusion, etc.), to reproduce exam reading by an experienced operator. RESULTS: The DL strategy displayed good diagnostic accuracy (88%), with an area under the curve (AUC) of 0.982. The precision (positive predictive value), recall score (sensitivity), and F1 score (a measure of test accuracy) were 83%, 95%, and 89% respectively. A ML algorithm considering all CMR features had a similar diagnostic yield to DL strategy (AUC 0.952 vs. 0.982; p = 0.39). CONCLUSIONS: A DL approach evaluating LGE acquisitions displayed a similar diagnostic performance for CA to a ML-based approach, which simulates CMR reading by experienced operators.