BACKGROUND
Thyroid nodules are very common, occurring in up to 50% of individuals. The concern about any nodule is whether it is a cancer. Fortunately, only ~5-6% of nodules are cancerous. Ultrasound is the most common technique used to distinguish which nodules require further evaluation with biopsy or surgery and which can be watched. Features of the ultrasound images can be used to determine the risk that the nodule is a cancer and decide which should be biopsied. Clinicians look for different features of the nodules that are concerning for cancer, such as darkness (hypoechoic), a mostly solid composition, a taller-than-wide shape, irregular margins, or the presence of calcifications within the nodule. None of these features are individually diagnostic and thus several systems, combining these features, are described to determine the risk of a given nodule. The one developed by the American College of Radiology (American College of Radiology Thyroid Imaging and Reporting Data System – ACR TI-RADS) is frequently used and has been shown to helpful to avoid unnecessary biopsies. However, even in this well-developed system, there is variation within individual readings and, even more so, between readers.

This study was performed to evaluate a specific artificial intelligence (AI) decision-support system (DSS), called Koios DS, to determine if this would improve the diagnostic accuracy and consistency among different readers.

THE FULL ARTICLE TITLE
Fernández Velasco P et al. Clinical evaluation of an artificial intelligence-based decision support system for the diagnosis and American College of Radiology Thyroid Imaging Reporting and Data System classification of thyroid nodules. Thyroid 2024;34(4):510-518; doi: 10.1089/thy.2023.0603. PMID: 38368560.

SUMMARY OF THE STUDY
This was a study of the ultrasound imaging of all nodules with cytological and/or histological results from a thyroid nodule clinic referral unit of a university hospital. It included all consecutive patients over 18 years of age with thyroid nodules and at least two ultrasound images with cytologic and/or histologic findings evaluated from June 2021 to December 2022.

The Koios DS uses ACR TI-RADS descriptors and scoring and gives an AI-derived adaptor that downgrades or upgrades the initial score. Six experienced clinicians were trained in the use of Koios DS and evaluated the ultrasound images from 172 patients twice, both with and without the AI adjustment. The classifications were compared between the two readings and to the patients’ histologic and/or cytologic findings.

A large number of nodules (81.3%) initially classified as ACR TI-RADS 3 (mildly suspicious) were reclassified as lower-risk. A quarter of those classified as ACR TI-RADS 4 (moderately suspicious) were also put into lower-risk categories. The AI-based DSS was better able to identify nodules that were suspicious that actually had suspicious cytology (improved from 14% to 16.1%) as well as those that were benign/not suspicious appearing and actually were benign (improved from 94.5% to 96.4%). In addition, the correlation between the observers improved significantly using AI-based DSS as compared to without. The AI system alone, without reader intervention, also showed good diagnostic performance.

WHAT ARE THE IMPLICATIONS OF THIS STUDY?
This study shows that AI has the potential to improve some of problems with ultrasound evaluations of thyroid nodules, thus improving diagnostic accuracy and consistency. It may also be helpful in supporting less experienced clinicians and improving the risk assessment process. Its limitations are that images were all obtained on a single ultrasound and read by two physicians with 20 years experience with only two images selected to be read by experienced physicians. In the real-world, thyroid ultrasounds are often obtained by technicians with varying experience, multiple images are evaluated by the reading physicians who also have varying experience. Thus, it would be important to continue to evaluate the system under different conditions. However, if in future studies, these results are confirmed, it represents the potential to minimize the number of nodules requiring biopsies while still capturing those patients who would benefit from biopsy and/or surgery.

— Marjorie Safran, MD

Table of Contents | PDF File for Saving and Printing