A Robust Ensemble Convolutional Neural Networks for Diagnosing Chest Diseases

Abstract

Radiologists employ X-ray images to differentiate various chest diseases. Given the intricate and meticulous nature of this diagnostic procedure, the assistance of automated models becomes imperative in detecting and diagnosing diseases from X-ray images. This research paper proposed a novel approach called Ensemble Convolutional Neural Network for Diagnosing Chest Diseases (ECDCNet), aimed at accurately and efficiently diagnosing fifteen different chest diseases through the analysis of X-ray images of the lungs. The ECDCNet model comprised a stack of five CNNs: ResNet152V2, DenseNet121, Inceptionv3, Vogg19, and Wavelet transform-CNN with various architectures and hyper-parameters to enhance the overall prediction performance. The proposed model applied the image segmentation for the lung's region using the U-Net model to localize and focus on the relevant space and facilitate the identification of specific radiological signs such as nodules, opacities, cavities, and consolidation. Furthermore, the study exploited three ensemble CNN strategies: Average voting, majority voting, and a proposed CNN-ensemble strategy called the Weighted Performance Metrics Ensemble Strategy (WPME) to set the weights of the prediction stage. The proposed WPME strategy used four evaluation measures for assessing the importance of each base CNN in the ensemble model, including precision, recall, F1-score, and accuracy, to enhance the prediction of the ensemble model. The proposed ECDCNet model achieved an accuracy of 95.3, 95.8 and 96.1% in the average voting, the majority voting, and the WPME strategy on a collected dataset of 110804 images for fifteen chest diseases. Further, it achieved an accuracy of 97.9, 98.2 and 98.9% in the average voting, the majority voting, and the WPME strategy on another public dataset of 13150 images for three chest diseases.