Optimizing Edge-AI for Medical Diagnostics: A Comparative Analysis of Custom CNNs vs. VGG16 Transfer Learning for Pneumonia Detection

Deep learning has revolutionized medical image analysis, yet deploying these models on low-resource devices remains a challenge. This study presents a comparative analysis between a resource-efficient Custom Convolutional Neural Network (CNN) and the industry-standard VGG16 architecture for detecting pneumonia in chest X-rays. Both models were trained on a dataset of 5,856 X-ray images. The VGG16 model achieved a higher overall accuracy of 93.0% compared to the Custom CNN’s 90.87%. However, the lightweight Custom CNN demonstrated superior sensitivity, achieving a Recall of 98% while requiring only 10MB of storage compared to VGG16’s ~528MB. These results suggest that custom lightweight architectures are the superior choice for mobile health applications in resource-constrained environments.