Utilization of Explainable Artificial Intelligence (XAI)-Powered Computer-Aided Detection (CAD) System on Chest X-Ray Abnormalities in Health Check-Ups

Abstract

Background: We designed a single-center retrospective study comparing the performance of commercially explainable artificial intelligence (XAI)-powered computer-aided detection (CAD) system of abnormal findings on chest X-rays (CXR) with that of non-experts, and pulmonology experts. Methods: A total of 1,262 images of 1,262 subjects (mean age 49 years; 52% female) and 1,252 images of 1,252 subjects (mean age 51 years; 51% female) were obtained from DICOM formats in Hakuai Hospital Health Check-up Center, in the pre-and post-implementing XAI-powered CAD period, respectively. The ultimate decision of abnormality on CXR was made by two pulmonology experts. The diagnostic accuracy metrics were measured accuracy and negative predictive value (NPV) for detecting abnormality on CXR. Results: XAI-powered CAD systems achieved an accuracy of 0.84 (95% confidential interval [CI] 0.82-0.86) and NPV of 1.00 (95% CI 0.99-1.00) to detect the abnormality on CXR. For determining nodular shadows, it was found to be non-inferior to the pulmonology experts with an accuracy of 0.94 (95% CI 0.92-0.95), and NPV of 1.00 (95% CI 0.99-1.00). It tended to overestimate the abnormality of heart enlargement and pleural thickening with a tendency for lower sensitivity. Conclusion: It seems likely that in the future, the most accurate screening CXR will be a double check combining with the pulmonology experts with XAI-powered CAD systems.

Ähnliche Arbeiten