Validation study demonstrates high classification accuracy for cardiac MRI scans, indicating potential for automated cardiovascular disease diagnosis.
Key Points
Develop and evaluate an explainable multi-scale deep learning framework (CardioFusion-XAI) to improve the accuracy and interpretability of four-class cardiac MRI disease classification.
Integrated WaveCLAHE-Net, which combines wavelet-based enhancement and contrast-limited adaptive histogram equalization, to preprocess and enhance image quality.
Constructed CardioXtract Fusion Net combining ResNet-50, EfficientNet-B0, and a Vision Transformer to capture local, multi-scale, and global features.
Evaluated performance on the Sunnybrook Cardiac Dataset and implemented Grad-CAM and t-distributed stochastic neighbour embedding for model interpretability.
CardioFusion-XAI achieved an overall classification accuracy of 98.89% and an area under the curve of 0.9998 on the Sunnybrook Cardiac Dataset.
The framework demonstrated a precision of 98.93% and an F1-score of 98.90% across the four cardiac disease classes.