Skip to main content
ENDO 2026 Main banner

Thyroid Biology and Disease

ORF40-04 - Aibx Software for Thyroid Cancer Diagnosis from Ultrasound Images: A Multicenter International Study

Monday, June 15, 2026
9:45 AM - 10:00 AM CT
Location: Room W179
Disclosure(s):
Rodis D. Paparodis, MD, FNLA: .NovoNordisk: Advisory Board (), Speakers' Bureau (); Ascendis Pharma: Advisory Board (), Grant Recipient (), Honorarium Recipient (); Sanofi: Advisory Board ()
Panagiotis Anagnostis, MD, MSc, PhD: No financial relationships to disclose

Introduction
AIBx is an Artificial Intelligence (AI) software for diagnosing thyroid cancer from ultrasound images. Initial studies demonstrated high sensitivity; the present study aimed to validate diagnostic accuracy across diverse ultrasound devices outside the original training data and to identify performance-influencing factors.

Methods
Two ultrasound images (sagittal and transverse) of thyroid nodules were retrospectively collected from patients undergoing fine-needle aspiration (FNA) or surgical resection at twelve endocrine clinics in three countries (United States, Greece, and Turkey). Corresponding FNA cytology (Bethesda categories II or VI) and/or surgical pathology reports were reviewed, and relevant clinical and laboratory data were recorded. AIBx software analyzed each image independently and classified nodules as benign (BEN) or malignant (MAL). Software output was compared with gold-standard cytology and/or pathology to calculate sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). A multivariate analysis assessed the impact of imaging, laboratory, and clinical parameters on the software’s diagnostic accuracy.

Results
A total of 998 consecutive patients were evaluated; 641 patients with complete imaging and diagnostic data were included in the final analysis, comprising 543 BEN and 98 MAL nodules. The mean patient age was 52.8±14.2 years, and 494 patients (77.1%) were female. Overall, AIBx achieved a sensitivity of 61.2% and a specificity of 55.2%. Concordant classification between sagittal and transverse images occurred in 463 cases (72.2%). In these concordant cases, sensitivity increased to 82.2% and specificity to 76.9%, with a PPV of 40.0% and an NPV of 95.8%. For sagittal images alone, sensitivity was 72.4%, specificity 70.0%, PPV 30.3%, and NPV 93.4%. For transverse images alone, sensitivity was 75.5%, specificity 68.7%, PPV 30.3%, and NPV 94.0%. When TIRADS classification was used as a prognosticator, sensitivity reached 100% in TIRADS 3 and 89.8% in TIRADS 5 nodules, but decreased to 46.2% in TIRADS 4 nodules. In a multivariate analysis assessing the predictive role of clinical, biochemical and imaging characteristics of the nodules, the software exhibited reduced diagnostic accuracy in nodules with ill-defined borders or mixed-pattern vascularity (peripheral and central simultaneously) and those in the TIRADS 4 category (p < 0.05 for all).

Conclusions
AIBx represents an innovative, user-friendly AI-based tool for differentiating benign from malignant thyroid nodules. This study demonstrates a high NPV and moderate diagnostic performance, particularly when sagittal and transverse image outputs are concordant. Although results are encouraging, algorithm optimization and validation are required before routine clinical implementation.