Efficient Pediatric Pneumonia Diagnosis Using Depthwise Separable Convolutions

Abstract

Pneumonia is the leading cause of death in children worldwide. A fast and accurate pneumonia diagnosis system can be helpful in saving a pediatric patient’s life and ensuring their long-term health. In recent years, A.I. research has attempted to develop reliable deep learning models for detecting pneumonia in chest X-ray images. The objective of this paper is to demonstrate that the use of depthwise separable convolutions provides an efficient pneumonia detection model. For this purpose, a novel 21-layer convolutional neural network, called PneumoniaNet, is presented. Most of the convolutional layers of PneumoniaNet use depthwise separable convolutions. Eight other customized pneumonia detection models, based on ImageNet pre-trained models, are also evaluated and compared with PneumoniaNet. PneumoniaNet is shown to be highly efficient without compromising effectiveness. In addition, the author demonstrates that the customized VGG16 has produced the highest test-set accuracy of 95.83%. In addition, and for completeness, PneumoniaNet’s robustness in case of "noisy" chest X-ray images is also analyzed.