Using Artificial Intelligence to Revise ACR TI-RADS Risk Stratification of Thyroid Nodules: Diagnostic Accuracy and Utility

Abstract

hyroid nodules are an extremely common finding at US and other imaging studies (1,2). Although most thyroid nodules are benign, many patients are subjected to a costly workup that may include one or more biopsies, follow-up imaging, and even diagnostic lobectomy (3). This contributes to the overdiagnosis of thyroid cancers that are not clinically significant (4). Over the past decade, multiple groups have developed biopsy guidelines for thyroid nodules based on their appearance at US, but some guidelines are difficult to apply and all lead to high false-positive rates (benign nodules for which biopsy is recommended). With these issues in mind, a committee of the American College of Radiology (ACR) created the Thyroid Imaging Reporting and Data System (TI-RADS) to determine if thyroid nodules depicted at US require biopsy or follow-up (5). Nodules are awarded points based on features in five categories-composition, echogenicity, shape, margin, and echogenic foci. The more suspicious the feature, the higher its point value. Points are summed to categorize a nodule into one of five TI-RADS risk levels, TR1 to TR5 (Table 1). Management recommendations are determined by using the risk level and the maximum size of the nodule. The points assigned to each feature in ACR TI-RADS were based on evidence in the literature and expert consensus. Therefore, it is possible that the performance of the system could be improved by optimization of the points assigned to each US feature. Given the problem of overdiagnosis in thyroid imaging, this might improve specificity without sacrificing sensitivity. Indeed, the ACR TI-RADS committee recognized that certain features may warrant higher or lower point values to achieve optimal performance (5).

A Comparative Analysis of Two Machine Learning-Based Diagnostic Patterns with Thyroid Imaging Reporting and Data System for Thyroid Nodules: Diagnostic Performance and Unnecessary Biopsy Rate

“…The diagnostic performance and the unnecessary biopsy rate of the ML-assisted US visual approach and the US radiomics approach were superior to the ACR TI-RADS evaluation in this study. The ACR TI-RADS evaluation results for predicting thyroid malignancy in this study (sensitivity, 85.4-90.9%; specificity, 45.2-52.5%) are similar to previous retrospective studies that used the ACR TI-RADS (sensitivity, 74.7-92%; specificity, 44-67.3%), and these results support the importance of ACR TI-RADS in the management of thyroid nodules (6)(7)(8). Several studies have attempted to apply ML techniques to process a series of collected US features from the radiologist's interpretation.…”

“…Some retrospective studies have shown encouraging results that improve the diagnostic accuracy and reduce the nodule biopsy recommendations in comparison to other biopsy recommendation guidelines. However, the risk stratification system of the ACR TI-RADS for thyroid nodules is relatively complicated to apply in clinical practice, and it has low diagnostic specificity (44-67.3%) (6)(7)(8). Due to the high prevalence of thyroid nodules and the low prevalence of biologically significant malignancy, a noninvasive and reliable diagnostic method with a high diagnostic sensitivity and specificity is highly desirable to identify which nodules warrant FNAB (9,10).…”

The ultrasound risk stratification systems for thyroid nodule have been evaluated against papillary carcinoma. A meta-analysis

“…The remaining 128 papers were retrieved in full-text and 9 studies were finally included in the systematic review ( Fig. 1) [22][23][24][25][26][27][28][29][30]. No additional study was retrieved from references of included studies.…”

Updates on the Management of Thyroid Cancer

“…Several novel therapeutic strategies, both single agents and combination therapies, are under various phases of clinical trials and are being utilized for different stages of thyroid cancer. Several studies have utilized artificial intelligence and machine learning and their findings have substantially added to the current knowledge of imaging, genetic, and molecular features of thyroid nodules and thyroid cancer [106][107][108][109][110]. The constant improvements in these technologies can give rise to more advanced meth-ods which may potentially revolutionize the diagnosis and management of thyroid cancer.…”