An artificial intelligence model demonstrated moderate performance for ruling out malignancy in incidental thyroid nodules detected on computed tomography, suggesting a potential role in reducing unnecessary downstream ultrasound imaging and biopsies, according to a research abstract presented at ENDO 2026, the Endocrine Society's annual meeting, in Chicago. 

Incidental thyroid nodules (ITNs) are commonly identified on computed tomography (CT) examinations and frequently trigger additional testing and surveillance. Researchers from Mayo Clinic and collaborating institutions developed and evaluated BioMedCLIP, a biomedical image encoder, designed to recognize CT-detected nodules unlikely to be malignant at the time of initial detection. 

The researchers retrospectively identified CT-detected ITNs from four Mayo Clinic sites between 2017 and 2023. Eligible patients had at least 12 months of follow-up and no previous thyroid imaging, thyroid disease, thyroid-related symptoms, or cancer.  Nodules were matched across CT examinations and reference standards according to size, location, and morphology. Malignancy status was determined by surgical histopathology or fine-needle aspiration Bethesda classification. For nodules lacking pathology, ultrasound-based Mayo Clinic risk stratification was used only when benign imaging features were present. Indeterminate nodules were excluded. 

The researchers manually segmented CT images using ITK-SNAP. Interannotator agreement was assessed using the Dice coefficient. BioMedCLIP was fine-tuned with this data and evaluated using five-fold cross-validation with an 80/20 train-test split. Model performance was measured by sensitivity, specificity, positive predictive value, negative predictive value, and overall accuracy. 

The final cohort included 381 patients, 62% of whom were women. Mean nodule size was 2.1 centimeters, 38% of nodules were malignant, and 60% of CT examinations were contrast-enhanced. Chest CT with and without intravenous contrast accounted for the largest proportion of examinations, representing 21% and 17%, respectively, followed by contrast-enhanced neck CT at 14%. 

Across the test sets, BioMedCLIP achieved a sensitivity of 79%,specificity of 49%,positive predictive value of 48%, negative predictive value of 79%, and overall accuracy of 60%. The researchers emphasized the model's negative predictive value as the feature most closely aligned with its intended clinical role of identifyingnodules unlikely to require further evaluation. 

The researchers noted that the study evaluated performance in a real-world clinical cohort rather than a curated imaging dataset, which they said better reflects routine practice. They also indicated that future work would expand the training dataset, optimize rule-out thresholds, and incorporate additional contextual information, such as radiology reports, to improve performance. 

David Toro Tobon, MD, reported serving as a consultant for Immunovant Inc. The other researchers had no financial relationships to disclose.

AACE Endocrine AI invited first author, Luis David Vilatuna, MD, of Mayo Clinic in Rochester, Minnesota, to elaborate on the study’s findings.  

Why does this study matter?  Recent studies have shown an increase in the incidental detection of thyroid nodules and the incidence of thyroid cancer over the past few decades. However, despite this rise in diagnoses, thyroid cancer mortality has remained largely unchanged. One of the most common imaging modalities for identifying incidental thyroid nodules is computed tomography (CT). 

The lack of standardized guidelines and reliable imaging markers to differentiate benign from malignant nodules detected on CT makes it challenging to determine which patients truly require further diagnostic evaluation. As a result, many patients undergo a cascade of additional investigations, often beginning with thyroid ultrasound and potentially leading to invasive procedures that may not be clinically necessary. This not only increases the physical and psychological burden on patients but also places unnecessary strain on healthcare systems. Our goal is to use AI to extract additional information from CT images and help clinicians make more precise decisions at the moment the nodule is first detected. By improving risk assessment early in the process, we hope to identify better which patients are most likely to benefit from further evaluation while reducing unnecessary downstream testing for others.   

What data stood out to you? What surprised us most was how variable the real-world documentation of incidental thyroid nodules can be. In many cases, the CT report did not describe the nodule in a standardized way, which makes sense because CT is usually performed for another clinical indication, not for dedicated thyroid evaluation. That was interesting for our study because we worked not only with images but also with text reports. The variability in how nodules were described became an important challenge when trying to incorporate report-based information into the model. It also reinforced the clinical problem we are trying to address. Once an incidental nodule is mentioned on CT and the patient is referred for dedicated ultrasound, that often opens the door to a cascade of follow-up imaging, surveillance, or biopsy. So, seeing the data directly made the overuse problem feel very concrete. 

How might the findings influence endocrinology clinical practice? This study could help clinicians more effectively select patients who truly need downstream evaluation after an incidental thyroid nodule is detected on CT. Our goal is not to replace radiologists’ clinical judgment, but to enhance it by using AI to extract additional imaging information that may not be readily visible to the human eye. 

Ultimately, we hope this approach can make the evaluation of incidental thyroid nodules more precise, reduce unnecessary testing and patient burden, and help ensure that downstream exams are directed toward the patients who need them most.