Ensemble localized patch residual convolution neural networks for pneumonia detection using chest X-ray

Background: Pneumonia is a life-threatening disease, especially in infants and older individuals. Chest X-ray is the most common method for the detection of lung inflammation especially when the causes of pneumonia are unknown. However, the evaluation of X-ray images requires special human expertise hence, the detection process can be error-prone and is also time- and resource-consuming. Automating the pneumonia-detection process from chest X-ray using a computer algorithm will improve precision, save time and resources. Methods: In this manuscript, a deep learning strategy called an ensemble of residual neural networks (eRes-NET), is presented for efficient pneumonia detection using X-ray images of the chest. The eRes-NET learns high-level features from multiple CNN models trained on smaller patches (25 × 25 pixels) of an image to identify distinguishing features of infection. In addition, to tackle the inadequate contrast of chest X-ray images leading to ambiguous diagnosis, the contrast-limited adaptive histogram equalization (CLAHE-DWT) technique is used to pre-process the images before pneumonia prediction. Results: Across a cohort of 8,266 chest X-ray images, eRes-NET achieved 93% accuracy and 94% AUROC on an independent posteroanterior test dataset. The approach performed favorably against 13 published methods (exceeding 8 under their reported protocols), noting differences in datasets and evaluation procedures. Conclusions: The proposed eRes-NET approach showed strong performance for pneumonia detection using chest X-ray images and may support efficient automated screening.