The diagnosis of heart disease using Cardiac Magnetic Resonance Imaging (MRI) is an essential area of medical diagnostics in the identification of various disorders such as coronary artery diseases, myocardial infarctions and heart failure. In recent developments in segmentation and classification networks, existing techniques are not accurate at the delineation of small, complex structures in MRI images. Frameworks such as U-Net are prone to overfitting, the merging of boundaries and poor integration of global–local features leading to poor detection of slight variations in heart structures. Vision Transformers (ViT) have been promising but are computationally intensive and not effectively process the multi-scale features in cardiac MRI images. To overcome these shortcomings, the new proposed framework of Hybrid Coyote-Driven Partitioning and Refinement (HCDPRA) for segmentation and Puma-Hybrid Vision Transformer (Puma-HViT) for classification. The HCDPRA method exploits the Coyote Optimization Algorithm (COA) to partition MRI images adaptively into cohesive regions and refine the boundaries through a dynamic exploration–exploitation strategy. This approach guarantees accurate segmentation through fine-grained local features (tissue boundaries) as well as global context (heart chambers). The Puma-HViT uses the Puma Optimizer to dynamically adjust learning rates and makes use of Hybrid Local-Global Attention (HLGA) to learn fine image features and long-range relationships. The Vision Transformer (ViT) is employed to represent global context by analyzing long-range relationships between patches in the image. The model obtains 94% Dice Coefficient for segmentation and 98.3% accuracy for classification on the benchmark MRI dataset. It provides improved results for benchmark Cleveland Heart Disease and lung cancer datasets. This method avoids overfitting, improves feature extraction and effectively models long-range dependencies for accurate heart disease prediction.
Amalraj et al. (Mon,) studied this question.