TI-RADS With Deep-Learning-Based Algorithm Guides Evaluation of Pediatric Thyroid Nodules
The American College of Radiology’s TI-RADS and a deep-learning (DL)-based algorithm trained on adult populations offer alternative strategies for evaluating thyroid nodules in children and younger adults, including guiding decisions to perform fine-needle aspiration, according to a manuscript in the American Journal of Roentgenology (AJR).
“Both ACR TI-RADS and the deep learning algorithm had higher sensitivity, albeit lower specificity, compared with radiologists’ overall impressions,” wrote co-first author Jichen Yang, BSE, from the department of electrical and computer engineering at Duke University in Durham, NC. Adding that the algorithm had similar sensitivity, but lower specificity, than ACR TI-RADS, “interobserver agreement was higher for ACR TI-RADS than for overall impressions,” Yang noted.
In the article, 39 patients (119 female, 20 male) aged ≤21 years with a thyroid nodule on ultrasound with definitive pathologic results from fine-needle aspiration and/or surgical excision were evaluated from January 1, 2004 to September 18, 2020. Single transverse and longitudinal images of one nodule per patient were then extracted. Three radiologists independently characterized nodules based on overall impression (benign vs. malignant) and ACR TI-RADS. A previously developed deep learning algorithm determined malignancy likelihood for each nodule, which was used to derive risk level.
Ultimately, for evaluating thyroid nodules via ultrasound in children and young adults, radiologists’ overall impressions—representing the current standard clinical approach—had mean sensitivity of 58.3% and mean specificity of 79.9%; ACR TI-RADS had mean sensitivity of 85.1% and mean specificity of 50.6%, and a deep learning algorithm had sensitivity of 87.5% and specificity of 36.1%.
“Given the heightened priority for sensitivity when evaluating thyroid nodules in children, compared with in adults, the findings support the continued exploration in children of ACR TI-RADS and of the deep learning algorithm,” the authors concluded.