Deep learning for the diagnosis of suspicious thyroid nodules based on multimodal ultrasound images

• DOI: 10.3389/fonc.2022.1012724
• Publication Date: 2022-11-08T13:06:36Z
• Authors: Yi Tao; Yanyan Yu; Tong Wu; Xiangli Xu; Quan Dai; Hanqing Kong; Lei Zhang; Weidong Yu; Xiaoping Leng; Weibao Qiu; Jiawei Tian
• Keywords: Thyroid Nodules; Kappa

This study aimed to differentially diagnose thyroid nodules (TNs) of Thyroid Imaging Reporting and Data System (TI-RADS) 3-5 categories using a deep learning (DL) model based on multimodal ultrasound (US) images explore its auxiliary role for radiologists with varying degrees experience.Preoperative US 1,138 TNs TI-RADS were randomly divided into training set (n = 728), validation 182), test 228) in 4:1:1.25 ratio. Grayscale (GSU), color Doppler flow imaging (CDFI), strain elastography (SE), region interest mask (Mask) acquired both transverse longitudinal sections, all which confirmed by pathology. In this study, fivefold cross-validation was used evaluate the performance proposed DL model. The diagnostic mature compared, whether could assist improving verified. Specificity, sensitivity, accuracy, positive predictive value, negative area under receiver operating characteristics curves (AUC) obtained.The AUCs differentiation 0.858 (GSU + SE), 0.909 CDFI), 0.906 CDFI 0.881 Mask), superior that 0.825-based single GSU (p 0.014, p< 0.001, p 0.002, respectively). highest AUC 0.928 achieved (G C E M)US, specificity 89.5% E)US, accuracy 86.2% sensitivity 86.9% M)US. With assistance, junior increased from 0.720 0.796 (p< 0.001), slightly higher than senior without assistance (0.796 vs. 0.794, > 0.05). Senior exhibited comparable (83.4% 78.9%, 0.041; 0.822 0.825, 0.512). However, significantly visual diagnosis 0.05).The models showed exceptional differential suspicious TNs, effectively efficacy TN evaluations radiologists, provided an objective assessment clinical surgical management phases follow.