Modified ACR TI-RADS and Modified AI TI-RADS for Thyroid Nodules: A Multicenter Retrospective Study

Background: Risk stratification systems for thyroid nodules are limited by low specificity. The fine-needle aspiration (FNA) biopsy size thresholds and stratification criteria are based on evidence from the literature and expert consensus. Our aim was to investigate the optimal FNA biopsy size thresholds in the American College of Radiology (ACR) Thyroid Imaging Reporting and Data System (TI-RADS) and artificial intelligence (AI) TI-RADS, and to revise the stratification criteria in AI TI-RADS. Methods: A total of 2596 thyroid nodules (in 2511 patients) on ultrasound examination with definite pathological diagnoses were retrospectively identified from January 2017 to September 2021 in six participating Chinese hospitals. The modified criteria for ACR TI-RADS were: (1) no FNA for TR3; (2) FNA threshold for TR4 increased to 2.5 cm. The modified criteria for AI TI-RADS were: (1) 6-point nodules upgraded to TR5; (2) no FNA for TR3; (3) FNA threshold for TR4 increased to 2.5 cm. The diagnostic performance and unnecessary FNA rate (UFR) of modified versions were compared with the original ACR TI-RADS. Results: Compared with original ACR TI-RADS, mACR (modified ACR) TI-RADS yielded higher specificity (73% vs 46%), accuracy (74% vs 51%), area under curve (AUC, 0.80 vs 0.70) and lower UFR (25% vs 48%; all P < 0.001) although the sensitivity was slightly decreased (87% vs 93%, P = 0.057). Compared with original ACR TI-RADS, mAI (modified AI) TI-RADS yielded higher specificity (73% vs 46%), accuracy (75% vs 51%), AUC (0.81 vs 0.70) and lower UFR (24% vs 48%; all P < 0.001), although the sensitivity tended to be slightly decreased (89% vs 93%, P = 0.13). There was no significant difference between mACR TI-RADS and mAI TI-RADS in the diagnostic performance and UFR (all P > 0.05). Conclusions: The revised FNA thresholds and stratification criteria of mACR TI-RADS and mAI TI-RADS may be associated with improvements in specificity and accuracy, without significantly sacrificing sensitivity for malignancy detection.