갑상선결절의 초음파 영상에서 인공지능을 이용한 병변 국소화

Abstract

Background and Objectives: The intersection over union (IoU) is a well-known indicator that measures how well the region of interest is delineated by artificial intelligence (AI). However, a definitive value for IoU threshold has not yet been established. Therefore, our aim was to establish a clear threshold value for IoU threshold and evaluate diagnostic performance across various thresholds. Materials and Methods: From January 2022 to January 2023, 218 thyroid nodules from 218 patients were included. IoU of AI’s ROI was calculated. Two radiologists evaluated whether the AI’s ROI is well delineated. Joinpoint analysis was used to examine the trend in radiologists scores according to IoU values. Generalized estimating equation was used to compare the diagnostic performance using cropped images by AI and a radiologist. Area under curve was evaluated using DeLong’s method. Results: The change point in the proportion of scores rated as 1, indicating appropriate inclusion by the AI, was found to be an IoU of 0.5 for one radiologist and 0.3 for the other. For IoU ≥0.5, 98.1% were rated as either 1 point by both radiologists or 1 point by one and 2 points by the other. But among 9 nodules with IoU less than 0.5, only one nodule was rated as I point. There were no statistically significant differences in diagnostic performance between radiologists and AI among the 209 nodules with an IoU ≥0.5. Conclusion: With an IoU threshold of 0.5, appropriate lesion localization was achieved, and there was no difference in diagnostic performance compared to the localization by an experienced radiologist.