DEVELOPMENT OF A CLINICAL DIAGNOSIS AND DECISION SUPPORT SYSTEM FOR CHEST RADIOGRAPHY USING CNN

Abstract

Chest X-ray (CXR) image is a widely used diagnostic tool for various chest diseases. Human interpretation of CXR images, however, has never been always effective. The diagnostic level of the radiologists along with other factors like cognitive ability, experience, fatigue, and other human-dependent factors may impair the diagnostic procedure with missed information, misinterpretation, and requiring more time and cost. Computer-aided analysis of CXR images has already demonstrated its potential over manual or human screening to facilitate rapid, correct, and low-cost diagnosis of chest diseases. Existing computer-aided systems are still not suitable for real-time applications due to limited findings, limited generalizability across wide datasets, and not being computationally and economically affordable as well. Therefore, an efficient Convolution Neural Network (CNN) based computer-aided decision support system, the CXRNet, was developed for the automatic detection of abnormalities from CXR images in a real-time clinical scenario. The proposed CXRNet model is a 16-layered CNN architecture with 5 output classes: Cardiomegaly, COVID, Normal, Pneumonia, and Tuberculosis. This architecture is trained with frontal CXR images obtained from various sources to improve the generalization of the model across multiple datasets. Upon testing the model on three different data distribution conditions (70% training and 30% testing, 80% training and 20% testing, and 90% training and 10% testing), it achieved a state-of-the-art performance with an average accuracy of 95.7%, a precision of 95.3%, a recall of 95.3%, and an f1-score of 95.3% for the multiclass classification task. The proposed CXRNet also demonstrates excellent performance on binary classification tasks with an average accuracy of over 98% for each disease condition. The results obtained from this work outperform several other custom-designed CNN architectures as well as pre-trained models-based architectures like ResNet, VGG, DenseNet, Xception, Inception, etc. Furthermore, with proper testing, validation, and debugging of the model in clinical practice, it can be successfully deployed as a decision support system for radiologists.