Cardiac amyloidosis: 119, LVH of other origins: 122
Fabry: 156 HCM: 59
Images
LVH: 462, normal apical 4 chamber (a4c): 1,807
HCM: 198 (genotype (+): 98, genotype (−): 100)
Cardiac amyloidosis: 119, LVH of other origins: 122
Fabry: 156, HCM: 59
AI method
Deep learning
Deep learning
Deep learning
Deep learning
Model design
Generative adversarial networks (GAN)
CNN (InceptionResnetV2)+RNN (LSTM)
CNN (VGG)
3D CNN (3D ResNet18)
Results
-score (0.83), accuracy (0.92), AUC (-)
-score (-), accuracy (0.84), AUC (0.84)
-score (-), accuracy (0.83), AUC (0.9)
-score (0.85), accuracy (0.91), AUC (0.91)
Strengths
The study proposed data-efficient deep learning for medical imaging, leveraging semisupervised GANs to enhance model performance by utilizing labeled and unlabeled data
The study assessed the deep learning model’s performance against established genotype prediction scores (Mayo Clinic I and II, Toronto), ensuring a reliable evaluation
The study demonstrated CNN’s performance against experienced human operators in visual analysis, assessing its potential to surpass traditional diagnostic methods
The study used a 3D ResNet18, well suited for the cardiac MRI data’s inherent 3D nature, effectively capturing spatial and temporal information
Weaknesses
The study recognizes the constraint of a small sample size, urging future research to enhance generalizability through larger and more diverse datasets
The study excludes patients with HCM phenocopies and poor image quality, which may introduce bias, limiting the model’s relevance to a broader population
The study was limited to a single institution and a small dataset of cine-CMR images. The generalizability of the deep learning model to other institutions and patient populations is still being determined
The study acknowledges a limitation in generalization capabilities due to the exclusive use of limited datasets from specific hospitals in Taiwan
HCM: hypertrophic cardiomyopathy; LVH: left ventricular hypertrophy.
