A computationally efficient hybrid ViT-CNN framework for automated cardiac MRI classification

Background and objective Automated cardiac magnetic resonance (CMR) image classification holds significant potential to enhance diagnostic precision and streamline clinical workflows. However, the prohibitive computational overhead of conventional deep learning architectures severely limits their seamless integration into standard hospital picture archiving and communication systems (PACS). To bridge this translational gap, this study introduces a resource-efficient hybrid deep learning framework that synergizes the strengths of Vision Transformers (ViTs) and Convolutional Neural Networks (CNNs) for automated CMR classification. Methods A specialized hybrid ViT-CNN architecture was engineered to maximize classification efficacy while strictly constraining computational latency. The proposed framework was trained and rigorously validated on a benchmark CMR imaging dataset for cardiac pathology categorization. The model's diagnostic performance was systematically evaluated utilizing standard quantitative metrics, specifically overall accuracy, F1-score, and the area under the receiver operating characteristic curve (AUC). Results The proposed framework demonstrated robust diagnostic capabilities, achieving an accuracy of 96.71%, an F1-score of 0.9671, and an AUC of 0.9960. Beyond high predictive precision, the optimized architectural footprint effectively mitigates prevailing computational bottlenecks, thereby validating its suitability for deployment on standard clinical hardware infrastructures. Conclusion The introduced hybrid ViT-CNN framework offers a highly effective and practical paradigm for automated CMR image analysis. By successfully reconciling diagnostic precision with computational efficiency, this methodology significantly advances the feasibility of deploying deep learning-based intelligent decision support systems in routine radiological environments.