Hybrid deep learning framework for lung disease diagnosis based on chest radiography images

Abstract

Early evaluation and diagnosis serve as the first crucial step towards effective patient management, especially in critical medical conditions like lung diseases. Computer-aided diagnostic systems can help radiologists make more precise diagnoses and reduce misinterpretations in lung disease diagnosis. Responding to the identified needs in the literature, this thesis introduces RVCNet, a hybrid deep neural network framework aimed at efficiently predicting lung diseases from a diverse X-ray dataset of multiple classes. The RVCNet is not developed in isolation but is ingeniously informed and improved by integrating the strengths of three deep learning techniques: ResNet101V2, VGG19, and a custom CNN model while limiting their weakness. This integration is careful and deliberate, designed to ensure that the framework capitalizes on the best features of each technique while minimizing their individual limitations. In the feature extraction phase of this new hybrid architecture, hyperparameter fine-tuning is used. In the subsequent classification phase, additional layers - including batch normalization, dropout, and dense layers - are incorporated to further refine the model’s predictive accuracy. The proposed method is applied to an award-winning publicly available dataset of COVID-19, non-COVID lung infections, viral pneumonia, and normal patients' chest radiography images. The experiments take into account 2262 training and 252 testing images. Results show that with the Nadam optimizer, the proposed algorithm has an overall classification accuracy of 91.27% with a sensitivity or recall of 98.30%, an AUC of 92.31%, precision of 90.48%, and an F1 score of 94.23% on the test data. Finally, these results are compared with some recent deep-learning models. For this four-class dataset, the proposed RVCNet has a classification accuracy of 91.27%, which is better than ResNet101V2 (81.35%), VGG19 (79.76%), VGG19 over CNN (79.05%), and other stand-alone models. The superiority of RVCNet serves as a testament to its robust design and its potential to be integrated into diagnostic procedures. These findings underscore the effectiveness of the proposed RVCNet in classifying lung diseases in various patient groups, thereby contributing a valuable tool to the ongoing efforts to diagnose and understand lung diseases, propelling the medical community a step closer to more accurate and early diagnoses