BACKGROUND
Thyroid nodules are very common in the general population. However, only 5-10% of all thyroid nodules are cancerous. Overall, thyroid cancer has an excellent prognosis. Many small (<1 cm) nodules are at very low risk and may not even need to be removed. Since thyroid nodules are commonly found, it can lead to overdiagnosis and overtreatment of thyroid cancer. Ultrasound is the best way to evaluate a nodule to determine the risk of cancer and whether a biopsy is needed. To avoid unnecessary thyroid nodule biopsy and increase accuracy for diagnosis of cancer, some methods have been developed to classify nodules based on the ultrasound findings, such as the Thyroid Imaging Reporting and Data System (TI-RADS) and the American Thyroid Association Risk Assessment.

Artificial intelligence (AI) is being tested out in a variety of medical conditions to assist in decision making. Deep learning is a subgroup of AI that could potentially have a more accurate or comparable diagnosis of thyroid nodules and cancer than experienced radiologists when used with important ultrasound characteristics. This multicenter study evaluated the use of a deep learning– based AI model to improve diagnosis of thyroid cancer by ultrasound images and compared the results with evaluation in clinical practice by physicians with different levels of experience.

THE FULL ARTICLE TITLE
Ha EJ et al 2023 Artificial intelligence model assisting thyroid nodule diagnosis and management: A multicenter diagnostic study. J Clin Endocrinol Metab. Epub 2023 Aug 25. PMID: 37622451.

SUMMARY OF THE STUDY
A total of 19,711 thyroid ultrasound images were obtained from 6163 consecutive patients with 7178 thyroid nodules from an academic hospital collected between July 2015 and May 2019. The inclusion criteria were being over 18 years of age with thyroid nodules ≥5 mm on ultrasound and available surgical specimens that were used to confirm if nodule was cancerous or noncancerous. Types of thyroid cancer included were papillary thyroid carcinomas, follicular carcinomas, and medullary carcinomas. A total of 17 different deep-learning algorithms were used and tested to differentiate cancerous and non-cancerous thyroid nodules. Two data sets from Ajou University Medical Center in Suwon, Korea were used, test set 1 from June to September 2015 and test set 2 from June 2020 and May 2021.

A diagnostic performance of deep-learning AI-based models achieved a sensitivity of 87% (the likelihood that a diagnosis of cancer is indeed cancer at surgery) and a specificity of 81.5% (the likelihood that a diagnosis of a benign nodule is indeed benign at surgery). In comparison, the average of 6 radiologists with different levels of expertise was a sensitivity of 82.3% and a specificity of 79.2%.

WHAT ARE THE IMPLICATIONS OF THIS STUDY?
These data suggest that a deep-learning AI algorithm of thyroid ultrasound images can improve accuracy of diagnosis for thyroid cancer and assist physicians with different levels of experience. Therefore, AI may be an important tool in the diagnosis of thyroid cancer in clinic practice, by providing accuracy and minimizing errors.

— Joanna Miragaya. MD

Table of Contents | PDF File for Saving and Printing