Evaluating Deep Learning Model Performance on Raspberry Pi 4 for COVID-19 Diagnosis

Abstract

The global coronavirus disease (COVID-19) pandemic has highlighted the urgent need for accessible, effective, and accurate diagnostic tools, especially in low- or no-service settings. Chest X-ray, a widely used diagnostic test, provides a rapid and non-invasive way to identify COVID-19 symptoms. Deploying deep learning models for COVID-19 detection on resource-constrained edge devices, such as the Raspberry Pi 4, requires a balance between model accuracy, inference speed, and hardware limitations. This study evaluates the performance of four deep learning models—ResNet18, ResNet50, DenseNet121, and SqueezeNet—based on key metrics such as inference time, memory usage, and model size after TensorFlow Lite conversion. Experimental results show that SqueezeNet offers the best trade-off, achieving the fastest inference time (10.76s per 100 images) and the lowest memory usage (2.8MB), making it the most suitable for real-time edge deployment. In contrast, ResNet50, despite its high accuracy, has the longest inference time (42.32s) and highest memory consumption (90MB), limiting its feasibility for Raspberry Pi-based applications. The findings highlight the importance of selecting lightweight architectures for efficient and scalable deep learning-based COVID-19 detection on edge devices. Detailed performance parameters including sensitivity, specificity, accuracy, reproducibility, and inference time were analysed to assess the suitability of each model for clinical trials. This study demonstrates the revolutionary potential of combining deep learning with portable devices to provide effective diagnostic tests for COVID-19 and similar respiratory diseases, address healthcare inequities, and facilitate timely interventions for epidemic control.