Experience-dependent integration of a commercial AI system with C-TIRADS for thyroid nodule assessment: a...

To evaluate the diagnostic performance of a commercial AI-assisted ultrasound system and C-TIRADS-based assessment among sonographers with different levels of experience, and to explore how two AI-assisted adjustment strategies performed in a retrospective clinical workflow for thyroid nodule evaluation.

This retrospective single-center study included adult patients with thyroid nodules who underwent conventional ultrasonography and commercial AI-assisted analysis at our hospital between December 2022 and December 2023 and had available reference-standard outcomes. Three senior and three junior sonographers independently reviewed stored ultrasound images and videos and reached group-consensus C-TIRADS classifications while blinded to AI results and final diagnoses. Two AI-assisted integration strategies were evaluated: Strategy 1 required a one-level adjustment of the initial C-TIRADS category according to AI output, whereas Strategy 2 allowed flexible reconsideration after review of the AI result. Surgical histopathology or cytological results with follow-up-supported benign classification were used as reference standards. Diagnostic performance, agreement with reference standards, and biopsy-related indicators were compared across approaches.

A total of 389 thyroid nodules from 343 patients were included. Because the cohort required cytological or pathological confirmation and complete imaging data, the study population was enriched for malignancy (58.10%). The AI system showed higher sensitivity and accuracy than the junior sonographer group and comparable performance to the senior sonographer group. Among senior sonographers, the flexible strategy was associated with improvements in sensitivity, accuracy, negative predictive value, agreement, and AUC, whereas the mandatory strategy did not yield significant overall benefit. Among junior sonographers, the mandatory strategy was associated with improvements in sensitivity, accuracy, negative predictive value, agreement, and AUC, whereas the flexible strategy showed limited benefit. In this selected cohort, the more favorable strategy for each experience group was also associated with lower unnecessary biopsy rates and higher positive biopsy yields, without a significant increase in missed diagnoses.

In this retrospective single-center, pathology-enriched cohort, the effect of AI-assisted C-TIRADS adjustment appeared to vary by sonographer experience. Mandatory adjustment may be more helpful for junior sonographers, whereas flexible integration may be more useful for senior sonographers. These findings should be interpreted as preliminary workflow observations and require validation in larger prospective multicenter studies.