BACKGROUND
Thyroid nodules are very common, affecting up to half of the US population. The concern about a thyroid nodule is whether it is a cancer. Thyroid ultrasound is the best initial study to determine whether a nodule is a cancer. The ultrasound characteristics, including the size, the appearance, the margins, blood flow and calcifications, all contribute to the determination of which thyroid nodules can be followed and which need a biopsy. This is known as risk-stratification. There are 2 major ultrasound-based algorithms currently being used to help risk-stratify the nodules: the American Thyroid Association (ATA) risk stratification system and the American College of Radiology Thyroid Image Reporting system (ACRTIRADS). These ultrasound algorithms identify high-risk nodules that require further assessment with a thyroid biopsy. An important consideration in clinical practice is the extent to which thyroid nodule risk stratification is consistent and reproducible across evaluators with variable degrees of expertise. This is particularly important, given that an inaccurate assigned ultrasound risk can lead to incorrect management recommendations.

The use of computer-aided diagnosis (CAD) systems based on machine learning has emerged as a possible solution standardize this process. Multiple CAD systems are available and are undergoing evaluation with the goal that they can support objective, accurate, and reproducible evaluation of thyroid cancer risk in practice and ultimately improve clinical outcomes for patients with thyroid nodules. An important part of this evaluation are thyroid nodules considered challenging by clinicians. This small study aimed to evaluate the diagnostic accuracy of a CAD system in the evaluation of thyroid nodules deemed difficult by clinicians.

THE FULL ARTICLE TITLE
Reverter JL et al 2022 Reliability of a computer-aided system in the evaluation of indeterminate ultrasound images of thyroid nodules. Eur Thyroid J 11(1):e210023. PMID: 34981749.

SUMMARY OF THE STUDY
This study evaluated thyroid nodules deemed difficult to interpret by a single expert senior endocrinologist with experience in ultrasound assessment. Records and ultrasound images of patients age >18 years who underwent total thyroidectomy or thyroid lobectomy based on thyroid biopsy results were included in the study. The ultrasound classification of nodules deemed “difficult to interpret” was defined as thyroid nodules showing patterns of mixed benign and concerning components. Subsequently, the thyroid nodules deemed difficult to interpret were analyzed by a CAD system, as well as by five human experts with expertise in ultrasound examination for the assessment of malignancy risk, based on the ACR TI-RADS classification. The five reviewers were unaware of the final diagnosis after surgery. Nodules were further classified as agreeing if there was agreement between ≥3 human observers and disagreeing if ≤2 observers agreed. The diagnostic performances of the readers versus the CAD system were compared.

There were 300 thyroid nodules considered, each with one ultrasound image, from which 80 were considered difficult to interpret by the senior endocrinologist. After assessment by the five human reviewers, 37 (46.25%) nodules were classified as agreeing and 43 (53.75%) disagreeing. When analyzing the nodules in the agreeing group, both the clinician observers and the CAD system obtained similar levels of accuracy, as compared with the known surgical diagnosis (74.5% vs. 77%). However, when analyzing disagreeing nodules, both the experts and the CAD system had a lower degree of diagnostic accuracy when compared to the evaluation of agreeing images (57.1% vs 77.0% for experts and 70% vs 74.2% for the CAD system).

WHAT ARE THE IMPLICATIONS OF THIS STUDY?
The accuracy of cancer risk by clinical observers and a CAD system on thyroid nodules deemed difficult were similar when there was general agreement with the human reviewers. However, the CAD system was more reliable than reviewers for challenging images where there was disagreement among the human reviewers. While these data are promising, further studies are needed to determine the role of CAD in assisting ultrasound interpretation.

— Alan P. Farwell, MD

Table of Contents | PDF File for Saving and Printing