Chest X-ray (CXR) is a fundamental diagnostic tool in detecting thoracic pathologies. However, the interpretation accuracy can vary significantly among radiologists, particularly those less experienced. Deep-learning-based automatic detection algorithms (DLAD) have emerged as a promising solution to augment diagnostic precision. This population-based, multi-reader study evaluates the performance of a DLAD (Carebot AI CXR) in detecting four major thoracic pathologies—Atelectasis (ATE), Consolidation (CON), Pulmonary Lesion (LES), and Pleural Effusion (EFF)—compared to the diagnostic accuracy of six junior radiologists in a real-world clinical setting. We retrospectively analyzed CXR images (n=999) from a mid-sized hospital, reflecting real-world prevalence of the studied findings. The DLAD’s performance was evaluated using sensitivity (Se), specificity (Sp), and likelihood ratios (PLR and NLR), and compared with radiologists’ assessments. A paired design was employed to compare Se and Sp with confidence intervals (CI) and p-values. The proposed DLAD demonstrated superior Se across all pathologies, with values of 0.938 (CI: 0.832–0.979) for ATE (n=48), 0.946 (0.852–0.981) for CON (n=55), 0.940 (0.887–0.969) for EFF (n=134), and 0.818 (0.680–0.905) for LES (n=44), where DLAD achieved lower Se than two assessed radiologists in multi-reader study (RAD 3 & RAD 5), but the differences were not statistically significant. However, it achieved lower Sp compared to junior radiologists in all findings, showing values of 0.914 (0.894–0.931), 0.803 (0.775–0.829), 0.875 (0.852–0.895), and 0.879 (0.854–0.900), respectively. The results highlight the potential of integrating DLAD into clinical practice as a decision-support tool for less experienced radiologists. The proposed DLAD has the ability to increase the sensitivity of radiologists’ reading.
.jpg)
