DSA-MobileNet: Resource-aware dynamic sparse attention for efficient cardiac MRI diagnosis on edge devices

Computational redundancy frequently hinders efficient cardiac magnetic resonance imaging (MRI) analysis, as standard deep learning architectures excessively process background noise and restrict deployment on resource-constrained edge devices. This study proposes DSA-MobileNet, a novel resource-aware framework integrating a MobileNetV3 backbone with a dynamic sparse attention (DSA) mechanism. During inference, the DSA module dynamically prunes non-salient tokens based on feature magnitude and real-time hardware constraints, concentrating the computational budget on high-entropy regions such as ventricular boundaries. The proposed model was evaluated using the publicly available Automated Cardiac Diagnosis Challenge (ACDC) dataset through a rigorous patient-wise -fold cross-validation protocol. Experimental results indicate that DSA-MobileNet attains an average classification accuracy of , an F1-score of , and a specificity of across the folds, significantly outperforming baseline architectures such as Hybrid Adaptive Fusion and Swin-CNN. These findings establish that intelligent feature pruning yields a scalable, high-fidelity solution for cardiac diagnostics on edge hardware without compromising analytical precision.