OBJECTIVE

To determine if the artificial intelligence-based Thyroid Imaging, Reporting and Data System (AI TIRADS) would perform better than the American College of Radiology (ACR) TIRADS in monitoring malignant thyroid nodules not recommended for biopsy using follow-up thresholds.

METHODS

A total of 3499 thyroid nodules with surgical histopathology and ultrasound features were retrospectively reviewed and categorized using ACR TIRADS and AI TIRADS. The recommendations for biopsy and follow-up divided nodules into three groups 1) fine needle aspiration (FNA), 2) follow-up ultrasound, and 3) no further evaluation.

RESULTS

Of the total 1608 malignant nodules in this study, 974 malignant nodules would not be biopsied in ACR TIRADS compared with 967 in AI TIRADS. While 60.0% (584/974) of these non-biopsied malignancies could be followed-up by ultrasound in ACR TIRADS and 62.8% (607/967) in AI TIRADS. For the malignancies of no further evaluation, 97.4% (380/390) were sized <10 mm in ACR TIRADS and 93.3% (336/360) in AI TIRADS. Compared with ACR TIRADS, AI TIRADS had lower unnecessary FNA rate and missing cancer rate (41.0% vs 47.8% and 22.8% vs 27.5%, P < .05, respectively) while having higher specificity and AUC as well as lower sensitivity (65.0% vs 57.9%, 0.895 vs 0.881, and 96.1% vs 97.8%, all P < .05).

CONCLUSIONS

Using the follow-up thresholds, more than half of the malignancies not being biopsied were monitored by ultrasound in both ACR TIRADS and AI TIRADS, and AI TIRADS had lower missing cancer rate. More than 90% of malignancies recommended for no further evaluation were <10 mm in diameter.