Optimizing ResNet50 for Medical Image Classification: A Comparative Study of Ghost Modules, Pruning, and Knowledge Distillation

Abstract

Computed Tomography (CT) imaging plays a vital role in assessing lung disease severity in COVID-19 patients. However, deploying deep learning models such as ResNet50 for automated severity classification (mild vs. severe) remains challenging in resource-constrained medical environments due to their high computational demands. This study presents a comparative analysis of three optimization techniques, namely ghost modules, pruning, and knowledge distillation, to enhance the efficiency of the ResNet50 model while maintaining high diagnostic accuracy. The optimized models were trained and evaluated using a real-world dataset comprising 817 CT images collected from public hospitals in Babylon Province, Iraq. Experimental work indicates that a GhostNet-augmented ResNet50 improved to a record 98.4% accuracy, a knowledge distillation-based variant achieved 93%, and a pruned ResNet50 variant achieved 91%. These results indicate a better balance between performance and computation achieved by ghost modules, supporting their particular relevance to real-time applications within resource-constrained medical systems.