An Improved Lightweight DenseNet-201 Model for Pneumonia Detection on Edge IoT

Pneumonia is an infection of the lungs that poses a significant threat to public health. An early diagnosis of pneumonia is crucial to detect possible progression to serious lung damage and initiate appropriate antibiotic treatments. Recently, deep learning (DL) has emerged as a promising tool for detecting pneumonia from medical images including chest X-rays. However, due to the extensive memory requirements, the realization of DL on limited memory-based edge devices such as mobile or Internet of Things (IoT) units is challenging. In this paper, we investigate the use of several DL methods including VGG-19, ResNet-50, MobileNet-V2, and DenseNet-201 for detecting pneumonia in chest X-rays. Further, we propose an improved version of the DenseNet-201 model that is able to achieve a significantly high 5-fold cross validation accuracy, sensitivity, specificity, and kappa score up to 97.1
^%
, 96.3%, 99.2% and 92.5% respectively with respect to the-state-of-the-art pre-trained models and the recent related work. Further, we exploit the use of linear post training quantization for optimizing the proposed modified DenseNet-201 and reducing the allocated memory for the model up to 4.2 times without compromising its performance. The proposed optimized model would eventually facilitate an accurate and real time pneumonia detection on edge IoT devices.