E2374. Artificial Intelligence Augmented Interpretation for Chest Radiographs: Real World Experience in Health Checkup Population
Authors
Kwang Nam Jin;
Seoul National University Boramae Hospital
Eun Young Kim;
Gil Medical Center, Gachon University College of Medicine
Young Jun Cho;
Konyang University Hospital School of Medicine
Objective:
We aimed to evaluate artificial intelligence (AI)-augmented interpretations of chest radiography (CXR) with grayscale softcopy presentation state (GSPS) display on picture archiving and communication systems (PACS) in a multicenter health checkup population.
Materials and Methods:
In this retrospective study, we analyzed 11,770 consecutive radiology reports from 3 healthcare centers, including 5778 (49.1%) reports with AI augmentation. A commercially available AI system was adopted on PACS with GSPS objects, which automatically displayed over their corresponding CXR images with an abnormality score for ten common radiological abnormalities (pneumothorax, mediastinal widening, pneumoperitoneum, nodule/mass, consolidation, pleural effusion, linear atelectasis, fibrosis, calcification, and cardiomegaly). The concordance rate between radiologists` interpretation and AI results was calculated. Concordance was achieved when the radiology reports were consistent with the AI results (“accept”), the radiology reports were partially consistent with the AI results (“edit”), or had additional lesions compared with the AI results (“add”). There was discordance when the AI results were rejected in the radiology report. In addition, the radiologists' reading times were compared between radiology reports with and without the AI system as a diagnostic aid.
Results:
Among 5778 participants (2,712 men; mean age, 52 years), thoracic abnormalities were found in 584 (10.1%) on CXR radiology reports and 1148 (19.9%) on AI results. The concordance rate was 86.6% (accept: 86.1%, edit: 0.3%, and add: 0.2%), and the discordance rate was 13.4%. Except for 1,529 cases (13.0%) for whom reading time data were unavailable or unreliable, the median reading time was slightly higher in reports with AI systems than without AI (median, 11-sec vs. 10-sec, p = 0.007).
Conclusion:
AI augmentation integrated on PACS and GSPS displays can be used as a reliable advisor for the radiologist's interpretation with minimal increase in reading time in the health check-up population. Integrating AI augmentation into the CXR screening workflow is usable in a real clinical environment.
