From Cloud to Edge: Evaluating CNNs for Lung Disease Classification for Resource-Constrained Settings

This study investigates the application of deep convolutional neural networks (CNNs) for lung disease classification from chest X-ray images, with a focus on deployment in resourceconstrained settings on low-power edge devices. Six state-of-the-art CNN architectures, including ResNet101, DenseNet201, MobileNetV3-Large, EfficientNetV2-B0, InceptionResNetV2, and Xception, were evaluated on a base dataset and an augmented dataset to assess their generalization capacity. Data augmentation showed varied impacts: ResNet101 exhibited significant improvement in validation accuracy with augmented data, indicating enhanced generalization, while models like InceptionResNetV2 showed limited improvement, suggesting underfitting. Additionally, inference performance on the Raspberry Pi 4 Model B demonstrated the impact of model complexity on speed and latency. MobileNetV3-Large and EfficientNetV2-B0 emerged as the most suitable models for real-time deployment, balancing accuracy with computational efficiency, whereas deeper models incurred higher latency, making them less practical for edge deployment. These findings underscore the importance of data augmentation, model selection, and inference efficiency in optimizing CNN-based lung disease classifiers for real-world healthcare applications. Future work will extend the evaluation to include heterogeneous edge devices, further exploring scalability and energy efficiency for real-time deployment in rural clinical settings with limited resources.