AI-Based Chest X-Ray Disease Detection and Clinical Recommendation System Using DL

Abstract

Chest radiography is one of the most common imaging modalities utilised for early diagnosis of pulmonary and thoracic abnormalities which permit early clinical intervention and care of the patient. It is cheap, non-invasive, available in most of the health care institutions and plays a significant role in diagnosis of many respiratory and cardiovascular illnesses. DL has optimized the efficiency, uniformity and usability of computer-aided diagnosis tools, along with the automated clinical recommendation systems, which enable doctors to make more accurate and faster suggestions. Diagnosis takes time, and is also highly dependent on experienced radiologists and suffers from inter-observer variability, particularly in a healthcare system with limited access to qualified radiologists. Based on TorchXRayVision, a public chest Xray dataset was used to recognise a few thoracic anomalies such as pneumonia, lung opacity, pneumothorax, cardiomegaly, pleural effusion, atelectasis, oedema, consolidation and other chest disorders. The images gathered were then converted to greyscale, scaled, normalised and transformed to a tensor for inference with DL model. The base CNN was DenseNet-121, which is a multi-layered CNN with transfer learning approach to enhance the feature extraction and classification performance and this was used for the multi-label disease prediction. The predicted results were then added to a clinical recommendation module which automatically provided decision support to patients and health professionals including disease description, symptoms, treatment recommendation, medicines, specialists, risk assessment, emergency preparedness, prevention and a downloadable PDF report. The performance of the evaluation was based on probability scores of disease, confidence band, multi-label prediction capacity, consistency prediction and accuracy of clinical suggestion. The DenseNet121 is stable and demonstrates good detection performance as the typical predictions are close to 65.77% confidence level in lung opacity and ranks the various chest anomalies according to the probability of them occurring. The proposed framework combines the accurate DL based diagnosis with the comprehensive clinical decision supporting features, resulting in significant reduction in the diagnostic burden and consequently improving the early detection of disease and accessibility of healthcare in the hospital as well as in the remote healthcare settings.