Machine learning model improves chest X-ray and radiograph image diagnoses
Machine learning model improves chest X-ray and radiograph image diagnoses lead image
Chest radiographs, or chest X-ray images (CXRs), are taken more than 2 billion times annually to diagnose lung issues such as pneumonia, tuberculosis, early-stage lung
cancer, and COVID-19. However, interpreting these images is complicated — it is thought that more than 22% of images are misinterpreted.
Magar et al. developed a convolutional neural network to automate and improve the analysis of CXR images for five common conditions: cardiomegaly, COVID-19, pneumonia, tuberculosis, and normal. The model, called CXRNet, was trained with publicly available images from hospitals and research centers. By creating CXRNet from scratch, the researchers ended up with a model with fewer trainable parameters, which helped overcome the limitations present in previous models.
“The model has achieved an accuracy of over 95%, equivalent or even superior to human-level performance,” said author Mohd Rashid. “Also, the comparative analysis of the model performance demonstrates the strength of the proposed CXRNet addressing the limitations of other research works.”
The authors are continuing to develop CXRNet for clinical practice with real-time validation, testing, and debugging. They are also working to classify more pulmonary diseases and expand their approach to evaluating CT and MRI images, which are often used as a follow-up to CXR imaging.
“We hope that our convolutional neural network model and the strategies implemented for network optimization in this work motivate the future research areas that can be improved for the further development and adaption of a robust decision support system,” Rashid said.
Source: “Development of a CNN-based decision support system for lung disease diagnosis using chest radiographs,” by B. T. Magar, M. A. Rahman, P. K. Saha, M. Ahma, M. A. Rashid, and H. Higa, AIP Advances (2025). The article can be accessed at https://doi.org/10.1063/5.0252595