Kotan R., Evren B.
Tomography, cilt.12, sa.10, ss.1-18, 2026 (SCI-Expanded, Scopus)
-
Yayın Türü:
Makale / Tam Makale
-
Cilt numarası:
12
Sayı:
10
-
Basım Tarihi:
2026
-
Doi Numarası:
10.3390/tomography12100144
-
Dergi Adı:
Tomography
-
Derginin Tarandığı İndeksler:
Scopus, Science Citation Index Expanded (SCI-EXPANDED), EMBASE, MEDLINE, Directory of Open Access Journals, PubMed
-
Sayfa Sayıları:
ss.1-18
-
İnönü Üniversitesi Adresli:
Evet
Özet
Background: Artificial intelligence (AI)-based ultrasound risk stratification has emerged as a promising tool for improving the evaluation of thyroid nodules. However, prospective evidence regarding its performance in routine clinical practice and its potential impact on biopsy recommendations remains limited. Methods: In this prospective single-center observational study, 1009 thyroid nodules undergoing ultrasound-guided fine-needle aspiration between May 2025 and May 2026 were independently evaluated using ChatGPT Plus as an investigational AI-assisted decision-support tool. Representative ultrasound images, including dimensional and Doppler images, were submitted to the AI without disclosure of the radiologist-assigned TI-RADS category or subsequent cytological or histopathological findings. AI-assisted and radiologist-assigned TI-RADS classifications were compared with Bethesda cytology, while postoperative histopathology served as the reference standard in the surgically treated subgroup. Diagnostic performance and discordance in risk classification between AI and routine radiologist assessment were evaluated. Results: The study included 1009 thyroid nodules from patients with a mean age of 51.5 ± 13.4 years, of whom 80.7% were female. Bethesda II was the most frequent cytological category (93.6%). AI classified 227 nodules (22.5%) as high risk (TR4–TR5), compared with 761 nodules (75.5%) by routine radiologist assessment (p < 0.001). Among Bethesda II nodules, AI assigned 172 (18.2%) to TR4–TR5 compared with 701 (74.2%) by radiologists. When ACR TI-RADS category-specific size thresholds were applied, FNA would have been recommended for 469 nodules (46.5%) according to AI-assisted classification and 993 nodules (98.4%) according to radiologist-assigned classification (p < 0.001), corresponding to 524 fewer modeled FNA recommendations with AI-assisted assessment. Histopathological evaluation was available for 73 surgically treated nodules, including 48 malignant and 25 benign lesions. Within this surgically verified subgroup, both AI-assisted and radiologist-assigned TI-RADS achieved a sensitivity of 100.0%, while specificity was 72.0% for AI-assisted assessment and 60.0% for radiologist assessment. Because histopathological verification was available only for surgically treated nodules, these diagnostic performance estimates apply specifically to the surgically verified subgroup and cannot exclude false-negative disease among nonoperated nodules. Conclusions: AI-assisted ACR TI-RADS assessment produced substantially fewer high-risk assignments than routine radiologist assessment, with the greatest divergence observed among cytologically benign nodules. These findings suggest a potential role for AI as a second-reader decision-support tool for refining thyroid nodule risk stratification and biopsy selection. However, because all nodules had already undergone FNA and histopathological verification was limited to surgically treated cases, the present study cannot establish the safety of AI-guided biopsy de-escalation or exclude false-negative disease among nonoperated nodules. Prospective management studies with longitudinal outcome verification are required to determine whether AI-assisted assessment can safely reduce unnecessary FNA.