A Multiclass Deep Learning Framework for Lung Diseases Classification with Chest X-ray Images

Abstract

One of the significant global health issues is lung diseases, and they are one of the top causes of morbidity and mortality around the world. Effective diagnosis is important to optimally treat patients, thus improving their prognosis. Chest X-rays, due to their cost, speed, and accessibility, are commonly used to identify abnormal pulmonary presentations. However, there are many overlapping anatomical structures, subtle signs of disease, and other similarities among different thoracic abnormalities, making manual examination of CXR images a challenge. To handle these challenges, this research proposes an innovative deep learning-based approach using chest X-ray images for multiclass classification of lung diseases. The proposed approach uses MobileNet+Adam optimizer in conjunction with a seven convolutional layer (CNN7) feature extraction block for enhanced learning of disease related features, thereby improving classification accuracy. A balanced dataset of 7,446 chest X-ray images consisting of three classes: pneumonia, normal, and pleural thickening was used to train and evaluate the proposed multiclass classification approach. Additionally, a wise comparison of seven different optimization algorithms (e.g., Adam, SGD, and RMSProp) was performed. The performance of the proposed multiclass classification approach was assessed using F1 score, accuracy, recall, precision, confusion matrix, and AUC. The results showed that the MobileNet+Adam optimizer (with 7 layers of convolution) produced the best overall performance, resulting in respective AUC values of 0.9473 for normal, 0.9996 for pneumonia, and 0.9474 for pleural thickening. Furthermore, comparison with previously published multiclass classification methods confirmed that the proposed deep learning multiclass classification approach outperformed baseline models and previous methods in the literature. The findings indicate that the proposed methodology using a MobileNet+Adam optimizer with 7 convolutional layers provides a robust, accurate, and computationally efficient solution for automated multiclass lung disease classification and has significant potential for incorporation into intelligent computer aided diagnostic systems for clinical decision support.