CardioFusion-XAI: a robust multi-scale and explainable framework for cardiac MRI-based disease classification

Abstract

Cardiovascular diseases remain a major global health burden, and early diagnosis using cardiac magnetic resonance imaging (MRI) is essential. However, manual interpretation is subjective, while existing deep learning models often show inconsistent preprocessing, limited multi-scale feature learning, and poor interpretability. CardioFusion-XAI, a hybrid explainable deep learning framework, was developed for four-class cardiac MRI classification. WaveCLAHE-Net integrates wavelet-based enhancement and contrast-limited adaptive histogram equalisation to improve image quality. CardioXtract Fusion Net combines ResNet-50, EfficientNet-B0, and a Vision Transformer to learn complementary local, multi-scale, and global features. An enhanced classification head performs prediction, while Grad-CAM and t-distributed stochastic neighbour embedding support model interpretability. Using the Sunnybrook Cardiac Dataset, CardioFusion-XAI achieved 98.89% accuracy, 98.93% precision, a 98.90% F1-score, and an area under the curve of 0.9998. CardioFusion-XAI provides accurate and interpretable cardiac MRI classification. Further multicentre validation is required before clinical implementation.