Comparison of Diagnostic Performance in Thyroid Nodules on US: Deep Convolutional Neural Network Models vs Endocrinologists With Various Experiences

Abstract Objectives: To diagnose thyroid cancer, ultrasonography is a primary tool, but diagnostic accuracy varies according to the proficiency of clinicians. The aim this study was compare performance between deep convolutional neural network (CNN) models and endocrinologist with various experiences. Methods: Patients who underwent fine needle aspiration at endocrinology department in Seoul National University Hospital, April 2014 June 2019, were reviewed. Among them, nodules which pathologically confirmed by surgery maximal diameter greater than 1cm included. Ultrasonography images reviewed 13 endocrinologists experiences: 0 month (E0, n=8), 1 year (E1, n=2), >5 years (E5, n=3). Results: Of total 451 nodules, 66.5% diagnosed as cancer 83.7% papillary (PTC). Sensitivity specificity CNN 85.3% 63.6%, respectively, its area under curve (AUC) 0.855. Compared CNN, mean E0 group significantly lower (Accuracy 68.7% vs 78.0%, P <0.001), after CNN-assistance, that improved (68.7% [before] 73.93% [after], = 0.008). E1 E5 groups showed similar CNN-assistance did not change it. Next, subgroup analysis performed histologic subtypes. AUC PTC (0.925) much higher other cancers including FTC (0.529). Interestingly, for only beginners (E0), also subset experienced (E1 E5). Conclusions: has good diagnosis PTC. Endocrinologist experience ultrasonography, beneficial improving especially