Thyroid nodules are risk-stratified by ultrasound and fine needle aspiration (FNA). However, about 15 to 30% of the nodules that undergo FNA may result in indeterminate result,1 corresponding to Bethesda categories III and IV, amongst which the malignancy rate is up to 40%.2 Traditionally, such indeterminate results have often led to diagnostic surgery, exposing patients to unnecessary operative risks and increased healthcare costs. In recent years, molecular markers have been used for better risk stratification. The introduction of the noninvasive follicular thyroid neoplasm with papillary-like nuclear features (NIFTP) classification has complicated the interpretation of molecular test results, notably affecting the positive predictive value (PPV). The newer generation of these tests have improved specificity, but PPV remains under 65%.3,4

The ThyroSeq panel tests genetic alterations associated with thyroid neoplasia and has undergone several iterations over the years. Its most recent iteration, ThyroSeq version 3(v3), assesses 112 genes using next-generation sequencing technology to identify point mutations, gene fusions, copy number alterations, and gene expression alterations.5 It classifies thyroid nodules as negative (or currently negative or negative but limited) or as positive for malignancy. It also gives a probability of malignancy ranging from low, likely low, intermediate, intermediate-high, and high.6 This can create diagnostic and management uncertainties for the treating clinician. Additionally, molecular testing is not without significant cost for the health system.7

Recent developments in artificial intelligence (AI), particularly deep learning (DL) algorithms applied to ultrasound images and whole slide images of FNA, present a promising complementary approach to molecular testing.8 The potential combination of molecular markers and AI-based image analysis could enhance diagnostic precision while potentially reducing healthcare costs associated with unnecessary surgeries. By combining imaging-based AI analysis with molecular marker results through algorithmic decision-making, diagnostic systems can potentially reduce diagnostic ambiguity, minimize inter-observer variability, and optimize patient management. Our study proposes such a hierarchical algorithmic approach for integrating AIBx V2 with ThyroSeq v3 molecular testing.

In this article, we determine the performance of artificial intelligence-based imaging (AIBx V2)9,10 model in indeterminate thyroid nodules. We also evaluate whether integrating AIBx V2 results with ThyroSeq v3 enhances diagnostic accuracy in the management of indeterminate thyroid nodules to decrease unnecessary surgery.