Lung Opacity Pneumonia Detection with Improved Residual Networks

Pneumonia detection is usually done by specialized highly trained specialists through the review of chest radiographs combined with vital signs, laboratory exams, and medical history. Though the presence of pneumonia in images appears in the form of areas of high opacity in the lungs which is difficult to differentiate from opaque areas caused by other conditions. The medical diagnosis of pneumonia is a very costly and time-consuming process. The motivation of this work is to automate the diagnosis process of pneumonia through image processing. In this paper, an automated pneumonia detection system is proposed using improved deep residual networks (ResNet) architectures, which are tested on the dataset of 30,227 DICOM Chest X-rays. Dataset was divided into 80% training and 20% testing, with 20% of the 80% used for training dedicated to validation. Two residual network models were used (Version 1 & Version 2), and results were also compared with three different CNN models as well as methods found in recent literature. The overall results indicate that the proposed ResNet (Version 2) method achieves higher accuracy than convolution neural networks and other recently proposed methods (Table 5). The proposed ResNet network of a depth of 110 and a batch size of 16 with epochs 80 achieved an average accuracy of 88.67%. An automated method is proposed and implemented in this work for the proper diagnosis of pneumonia using images of the DICOM chest X-rays dataset. The proposed method in this paper outperforms other methods from recent literature.