Abstract
AbstractObjectiveTo investigate the difference between an artificial intelligence (AI) system, fine-needle aspiration (FNA) cytopathology, BRAFV600E mutation analysis and combined method of the latter two in thyroid nodule diagnosis.MethodsUltrasound images of 490 thyroid nodules (378 patients) with postsurgical pathology or twice of consistent combined FNA examination outcomes with a half-year interval, which were considered as gold standard, were collected and analyzed. The diagnostic efficacies of an AI diagnostic system and FNA-based methods were evaluated in terms of sensitivity, specificity, accuracy, κ coefficient compared to the gold standard.ResultsThe malignancy threshold of 0.53 for an AI system was selected according to the optimization of Youden index based on a retrospective cohort of 346 nodules and then applied for a prospective cohort of 144 nodules. The combined method of FNA cytopathology according to Bethesda risk stratification system and BRAFV600E mutation analysis showed no significant difference in comparison with the AI diagnostic system in accuracy for both the retrospective and prospective cohort in our single center study. Besides, for the 33 indeterministic Bethesda system category III and IV nodules included in our study, the AI system showed no significant difference in comparison with the BRAFV600E mutation analysis.ConclusionThe evaluated AI diagnostic system showed similar diagnostic performance to FNA cytopathology combined with and BRAFV600E mutation analysis. Given its advantages in ease of operation, time efficiency, and noninvasiveness for thyroid nodule screening as well as the wide availability of ultrasonography, it can be widely applied in all levels of hospitals and clinics to assist radiologists for thyroid nodule diagnosis and is expected to reduce the need for relatively invasive FNA biopsies and thereby reducing the associated risks and side effects as well as to shorten the diagnostic time.
Publisher
Cold Spring Harbor Laboratory