Independent evaluation of the accuracy of 5 artificial intelligence software for detecting lung nodules on chest X-rays.

Abstract

Background: The integration of artificial intelligence (AI) into medicine is growing, with some experts predicting its standalone use soon. However, skepticism remains due to limited positive outcomes from independent validations. This research evaluates AI software's effectiveness in analyzing chest X-rays (CXR) to identify lung nodules, a possible lung cancer indicator.

Methods: This retrospective study analyzed 7,670,212 record pairs from radiological exams conducted between 2020 and 2022 during the Moscow Computer Vision Experiment, focusing on CXR and computed tomography (CT) scans. All images were acquired during clinical routine. The final dataset comprised 100 CXR images (50 with lung nodules, 50 without), selected consecutively and based on inclusion and exclusion criteria, to evaluate the performance of all five AI-based solutions, participating in the Moscow Computer Vision Experiment and analyzing CXR. The evaluation was performed in 3 stages. In the first stage, the probability of a nodule in the lung obtained from AI services was compared with the Ground Truth (1-there is a nodule, 0-there is no nodule). In the second stage, 3 radiologists evaluated the segmentation of nodules performed by the AI services (1-nodule correctly segmented, 0-nodule incorrectly segmented or not segmented at all). In the third stage, the same radiologists additionally evaluated the classification of the nodules (1-nodule correctly segmented and classified, 0-all other cases). The results obtained in stages 2 and 3 were compared with Ground Truth, which was common to all three stages. For each stage, diagnostic accuracy metrics were calculated for each AI service.

Results: Three software solutions (Celsus, Lunit INSIGHT CXR, and qXR) demonstrated diagnostic metrics that matched or surpassed the vendor specifications, and achieved the highest area under the receiver operating characteristic curve (AUC) of 0.956 [95% confidence interval (CI): 0.918 to 0.994]. However, when evaluated by three radiologists for accurate nodule segmentation and classification, all solutions performed below the vendor-declared metrics, with the highest AUC reaching 0.812 (95% CI: 0.744 to 0.879). Meanwhile, all AI services demonstrated 100% specificity at stages 2 and 3 of the study.

Conclusions: To ensure the reliability and applicability of AI-based software, it is crucial to validate performance metrics using high-quality datasets and engage radiologists in the evaluation process. Developers are recommended to improve the accuracy of the underlying models before allowing the standalone use of the software for lung nodule detection. The dataset created during the study may be accessed at https://mosmed.ai/datasets/mosmeddatargogksnalichiemiotsutstviemlegochnihuzlovtipvii/.