Accurate diagnosis of cardiac diseases from Magnetic Resonance Imaging (MRI) remains a challenging task due to the need for high predictive performance while maintaining computational efficiency. This paper presents a computing-centric and efficiency-aware deep learning framework for cardiac MRI analysis, explicitly designed for diagnostic and clinical research. The proposed model is based on a two-layer GhostNet architecture enhanced with a combined channel-attention mechanism that integrates squeeze-and-excitation (SE) operations, enabling effective recalibration of feature channels with minimal parameter overhead. Network training is performed using an improved Parrot Optimisation Algorithm, which is algorithmically and mathematically formulated to adaptively optimise model parameters, resulting in improved convergence behaviour and generalisation capability. To further reduce computational complexity, structured pruning is systematically applied and quantitatively analysed, with explicit evaluation of parameter count, floating-point operations (FLOPs), memory footprint, and accuracy trade-offs. This pruning strategy enables the development of a compact diagnostic model suitable for deployment on resource-constrained computing platforms without substantial degradation in predictive performance. The framework is evaluated on benchmark cardiac MRI datasets using accuracy, precision, recall, and F1-score, alongside comprehensive computing-efficiency metrics. Experimental results demonstrate a favourable accuracy–complexity balance and significant reductions in model size and computational cost compared to conventional convolutional neural network architectures.