Abstract X-ray is a primary care diagnosis method for lung diseases, which allows physicians to evaluate and plan the treatment course. X-ray machines are quite prevalent, making them a potent tool for analyzing new respiratory diseases. Therefore, various automated deep learning (DL)-based models have been developed to improve the diagnosis of lung diseases and treatment quality to ensure a better patient survival rate. In this work, a multi-stage classification approach is proposed to optimize the diagnostic accuracy by leveraging the strengths of different convolutional neural networks (CNNs). SqueezeNet, ResNet-50, and EfficientNet-b0 were selected due to their efficiency, high accuracy, and prior success in medical imaging tasks. Each CNN model was optimized with respect to batch size and optimizer type (adaptive momentum (Adam), root mean square propagation (RMSprop), and standard gradient descent with momentum (SGDM)) to achieve the best possible accuracy in chest X-ray analysis. The study utilizes the COVID-QU-Ex dataset and the Tuberculosis (TB) Chest X-ray Database. The classification was implemented in three stages: (i) SqueezeNet (accuracy = 99%) distinguishes Normal vs. Abnormal X-rays, (ii) ResNet-50 (accuracy = 98%) classifies Abnormal X-rays into COVID-19 or non-COVID categories, and (iii) EfficientNet-b0 (accuracy = 97%) further categorizes non-COVID cases into pneumonia (bacterial/viral) or tuberculosis. Performance metrics, including accuracy, F1-score, precision, receiver operating characteristics (ROC) curve, and specificity, were evaluated using the confusion matrix. To facilitate automated diagnosis, a MATLAB-based toolchain named ‘CovidApp’ has been developed for real-time classification of chest X-ray images. The performance of the developed CNN toolchain was tested against a new set of sample lung X-ray images obtained from literature. No additional preprocessing techniques were applied to the images. While the tool has not yet been tested in resource-limited settings, its implementation could aid early disease detection in regions with limited access to expert radiologists. Future work will focus on real-world clinical validation and enhancing the robustness of the DL models to ensure reliable deployment in diverse environments. Graphical abstract
Deepak et al. (Fri,) studied this question.