AI for Differentiation of Low vs High Risk Thyroid Nodules on Ultrasound

(ABSTRACT) Thyroid cancer is common, but up to 77% of thyroid cancers have been attributed to overdiagnosis. The problem of overdiagnosis is rooted in current management strategies for thyroid nodules that do not appropriately account for the fact that most thyroid cancers are indolent and do not impact morbidity or mortality. As a result, patients are exposed to unnecessary diagnostic workup and treatment, resulting excessive costs to both patients and healthcare systems. The workup for thyroid nodules begins with ultrasound (US), during which US images are analyzed to decide if a nodule looks suspicious and requires biopsy to assess for possible cancer. While several standardized criteria exist to guide this analysis and decision-making, they suffer from multiple problems including (1) low specificity, leading to a large number of benign biopsies, (2) high interreader variability, causing heterogeneity in decision-making, and (3) an emphasis on a binary malignancy diagnosis rather than an assessment of nodule risk in terms of morbidity or mortality. With these factors in mind, our goal is to foundationally change how nodules are analyzed on US and to revise current systems that contribute to overdiagnosis. To accomplish this, our research will proceed with three specific aims. First, we will establish a definition of high- and low-risk nodules that focuses on mortality rather than cancer status. Many types of thyroid cancer can be watched rather than removed, and these cancers will be considered low-risk. Using these innovative definitions, we will establish a comprehensive multi-institutional repository of thyroid nodules that will allow us to better understand the previously unknown incidence of high-risk nodules and rates of unnecessary biopsies. In the second aim, we will develop a novel multi-task deep learning model for predicting high- vs low-risk thyroid nodules on US. An accurate model that can consistently diagnose low-risk nodules will reduce unnecessary biopsies and result in downstream cost savings. In the final aim, we will create a feature-based risk stratification system (RSS), which will also differentiate high- from lowrisk nodules. This system will mimic feature-based RSSs already used worldwide (but which focus on the less relevant question of benign vs malignant). We will compare the performance of our two systems (deep learning and feature-based) and demonstrate the potential for biopsy reduction. Ultimately, our goal is to accurately characterize low-risk nodules and demonstrate how focusing on risk rather than malignancy status could reduce unnecessary biopsies and surgeries. This project will be conducted by an interdisciplinary team including several members who have served in international leadership roles to help set standards for thyroid nodule management.