Artificial Intelligence-Based Automated Detection of Chest X-ray Abnormalities as Support for Young Radiologists.
Authors
Affiliations (4)
Affiliations (4)
- Policlinico Umberto I Hospital, Department of Radiological and Oncological Sciences and Pathological Anatomy, Sapienza University, 00161 Rome, Italy.
- Policlinico Umberto I Hospital, Department of Translational And Precision Medicine, Sapienza University, 00161 Rome, Italy.
- Department of Experimental and Clinical Biomedical Sciences, Radiodiagnostic Unit n. 2, University of Florence-Azienda OspedalieroUniversitaria Careggi, Largo Brambilla 3, 50134, Florence, Italy.
- Poliambulatorio Montezemolo, 00195 Rome, Italy.
Abstract
Chest X-ray (CXR) still represents the most performed radiological examination, but its interpretation may differ among readers. In the last years, several automatic detection tools have been developed to assist radiologists in CXR interpretation. We aimed to evaluate how artificial intelligence (AI)-based software may increase the performance of young radiologists in CXR interpretation compared to that of an experienced CXR radiologist. 500 CXRs were selected to generate a well-balanced dataset. For each patient, the CXR reading was conducted by an AI-based software (Lunit INSIGHT CXR) and, independently, by two general radiologists with less than one year of experience. Furthermore, after three months, the radiologists reviewed all the examinations with the assistance of the software. The following CXR findings were searched: atelectasis, calcification, cardiomegaly, consolidation, fibrosis, nodules, mediastinal widening, pleural effusion, pneumoperitoneum, and pneumothorax. Sensitivity and specificity were computed compared to an expert reader allowed to use AI assistance. Chi-squared tests were used to compare the readings. A total of 600 findings were identified by the senior radiologist. The sensitivity of young radiologists without AI was significantly lower than that of AI (p < 0.001), while the specificity was significantly higher for both (p < 0.001). With the assistance of AI, overall sensitivity increased in both radiologists (p < 0.001), while specificity decreased (p ≤ 0.004). AI software achieves very high sensitivity, minimizing the number of false negatives as much as possible. As a counterpart, there could be a nonnegligible risk of overdiagnosis. AI-based software assistance can yield a potential enhancement in sensitivity for young radiologists in the interpretation of CXRs. However, there might be a reduction in specificity.