Deep learning to diagnose cardiac amyloidosis from CMR findings. (25th November 2020)

• Record Type: Journal Article
• Authors: Aimo, A; Martini, N; Barison, A; Della Latta, D; Ripoli, A; Chiappino, S; Chiappino, D; Passino, C; Emdin, M
• Is Part Of: European heart journal. Volume 41:(2020)Supplement 2
• Publication Date: 2020-11-25
• Subjects: Cardiac Magnetic Resonance: Myocardium; Cardiology -- Periodicals; Heart -- Diseases -- Periodicals; 616.12005
• DOI: 10.1093/ehjci/ehaa946.0211
• Languages: English
• ISSNs: 0195-668X

Abstract

Abstract: Background: Cardiac magnetic resonance (CMR) is an important diagnostic technique for cardiac amyloidosis (CA). A deep learning (DL) approach to define the likelihood of CA based on automated interpretation of CMR images has never been attempted so far. Methods: 187 subjects underwent standard 1.5 T CMR examination (GE-Healthcare, Milwaukee, USA) as part of a diagnostic workup for either unexplained left ventricular hypertrophy or blood dyscrasia with suspected light-chain (AL) amyloidosis. Patients were randomly assigned to 3 subgroups, which were used for training (n=121, 65%), internal validation (n=28, 15%), and model testing (n=38, 20%). LGE images in different orientations (short-axis, 2- and 4-chambers) were selected as the most informative CMR features. A deep convolutional neural network was trained to classify CMR examinations as "amyloidosis" (probability ≥50%) or "no amyloidosis" (probability <50%) based on these features. Different learning strategies (data augmentation, batch normalization in convolutional layers, dropout before dense layers) were adopted to prevent model overfitting. Binary cross entropy was used as loss function. For comparison, a machine learning (ML) model based on gradient boosting trees was built for the binary classification of patients (amyloidosis vs no amyloidosis) based on clinical and imaging features extracted from the CMR exam. Results: CA was diagnosed in 101 subjects (54%; 45 AL, 56 transthyretin amyloidosis). A model including 2C, 4C and SA LGE images was created. In the test cohort, it allowed to diagnose CA with good diagnostic accuracy (84.2%), and an area under the curve (AUC) of 0.96 (Figure). The precision (positive predictive value), recall score (sensitivity), and F1 score (a measure of test accuracy) were 0.78, 0.94, and 0.86, respectively. An ML algorithm considering all available parameters (LV volumes and function, LGE presence and pattern, early darkening, pericardial and pleural effusion, etc.) displayed a similar diagnostic performance than the DL method (AUC 0.93 vs. 0.96; p=0.45). Conclusions: The deep learning technique allowed to create an accurate diagnostic tool for CA based on LGE patterns, which could be easily converted into an online platform for automated image analysis. Funding Acknowledgement: Type of funding source: None