What the study evaluated

The study evaluated the performance of Carebot AI CXR, a deep learning–based detection system, for identifying seven common radiological findings on chest X-rays in routine clinical practice. AI performance was compared with six radiologists of varying experience in a multi-reader design using 956 consecutive CXRs from a real hospital workflow, with expert consensus as ground truth.

Study results in clinical practice

Across all evaluated findings, AI consistently achieved high sensitivity, particularly for low-prevalence and clinically relevant abnormalities such as pulmonary lesions and pneumothorax. This significantly reduced false negatives compared to junior and intermediate radiologists. Specificity was generally lower than that of experienced radiologists, resulting in more false-positive alerts. In clinical practice, this confirms AI’s role as a safety-focused decision support tool that helps standardise detection quality and mitigates experience-related variability.

Key numbers

• CXRs analysed: 956

• Findings detected: 7 (atelectasis, consolidation, pleural effusion, pulmonary lesion, subcutaneous emphysema, cardiomegaly, pneumothorax)

• Pulmonary lesion sensitivity (AI): 90.5%

• Pulmonary lesion sensitivity (radiologists): 23.8–66.7%

• Pleural effusion sensitivity (AI): 95.3%

• Overall trend: higher sensitivity than most radiologists, lower specificity

• Greatest benefit: junior and mid-level radiologists