SUN-347 Factors Affecting the Diagnostic Accuracy of AIBx for Thyroid Cancer. A Prespecified Subgroup Analysis from an International Multicentric Study

Abstract Disclosure: R.D. Paparodis: None. A. Rizoulis: None. P. Anagnostis: None. J.C. Jaume: None. D.P. Askitis: None. D. Zianni: None. I. Androulakis: None. N. Angelopoulos: None. S. Livadas: None. A. Boniakos: None. D.D. Bantouna: None. Y.S. Guerra: None. K. Bayrakdar: None. G. Simeakis: None. A. Melhim: None. J. Thomas: None. INTRODUCTION The diagnosis of thyroid cancer using artificial intelligence (AI) represents a significant advancement. AIBx software, designed to analyze two thyroid ultrasound images per nodule, has demonstrated high negative predictive value in external validation studies. However, to date, no studies have investigated the underlying reasons for its failure in accurately diagnosing certain thyroid nodules. This study aims to identify the factors contributing to cases where AIBx produces incorrect diagnoses. METHODS We collected two ultrasound images from thyroid nodules (sagittal:S and transverse:T) undergoing FNA or surgical resection, in nine Endocrine clinics in Greece and two in the US, over 2 months, along with clinical data (age, gender, thyroid disease history and treatment), labs (TSH, TPO, Tg-Abs), FNA cytology and surgical histology. We analyzed the nodules images with AIBX and obtained a diagnosis of benign (BEN) or malignant (MAL). RESULTS 672 consecutive nodules were evaluated and the images analyzed; n= 495 were included in the present study, females n=381 (77.0%), BEN n=449, MAL n=46 (7.1%), age 53.5±14.7, BMI 26.9±5.3Kg/m2, TSH 1.69±1.22IU/L, TPO 92±366IU/L, Tg-Abs 77±291IU/L. The output was discordant between the S and T images in 134 cases (27.1%). We constructed three logistic regression models (LRM) to identify clinical (age, gender, BMI, TSH, TPO, Tg-Abs, treatment with thyroxine history), sonographic (nodule size, shape, echogenicity, margins, tall>wide, echogenic foci) and user-related factors (sonographer, device, year) leading to this discrepancy: the clinical parameters and user-related parameters did not predict that discordance (p>0.05 for all); the presence of irregular margins and macrocalcifications were significant though. S images failed to diagnose correctly 30.2% and T images did so in 32.4% of cases. A similar LRM found that the presence of irregular borders and low Tg-Abs (<30IU/L) enhanced that risk in S view, while only the presence of irregular borders was significant in T view. CONCLUSIONS AIBx is one of the AI models attempting to discern benign from malignant nodules based on an easy-to-use process evaluating two ultrasound images alone. In this international, multicentric study, discrepancy between these two images and inaccuracy appear related to specific sonographic features of the nodules studied. Further tuning of the software in that direction should allow to eliminate these and enhance its diagnostic accuracy. Presentation: Sunday, July 13, 2025