Artificial intelligence in ultrasound of thyroid nodules, prognosis of I-131 uptake

Abstract

BACKGROUND: Thyroid nodules are a prevalent issue, with an estimated incidence of 19% to 35% based on ultrasound examination and 8% to 65% based on autopsy findings [1]. In some cases, Plummer’s disease is observed, and nodular masses may be observed in 10% to 35% of Graves’ disease cases, with iodine accumulation of a different nature [2, 3]. One of the principal treatments for Graves’ and Plummer’s diseases is radioiodine therapy, which serves to exclude the possibility of malignancy in nodules. Furthermore, the pharmacokinetics of iodine is investigated, which represents the most time-consuming and labor-intensive stage of preparation for radioiodine therapy. In clinical practice, ultrasound is performed in accordance with the TI-RADS system, followed (if necessary) by fine-needle aspiration puncture biopsy, stratified according to the Bethesda system. However, the interpretation of ultrasound examinations is inherently subjective, whereas the use of decision support systems can reduce the number of fine-needle aspiration puncture biopsies by 27% and the number of missed malignant neoplasms by 1.9%. Furthermore, the quantitative characterization of nodal ultrasound may enhance the investigation of the pharmacokinetics of I-131 [4, 5]. AIM: The study aimed to develop a method for quantitatively characterizing ultrasound images of thyroid nodular masses for predicting malignancy and I-131 accumulation by nodular masses. MATERIALS AND METHODS: The study included 125 nodules with pathomorphologic findings (65 benign, 60 malignant) and 25 benign nodules (established by cytologic examination) of patients who underwent radioiodotherapy as part of the Russian Science Foundation grant project No. 22-15-00135. Longitudinal and transverse projections of thyroid nodules were obtained using GE Voluson E8 (36% of all benign nodules and 27% of malignant nodules) and GE Logiq E (64% of benign and 73% of malignant nodules). A pharmacokinetics study was conducted on 25 nodes obtained on a GE Logiq V2 device. The accumulation index of I-131 was determined after 24 hours. A spatial adjacency matrix, gray level line length matrix, gray level zone size matrix, and histogram were employed to investigate features based on ultrasound images. RESULTS: The malignancy prediction model, developed on the basis of the most significant features and after KNN correlation analysis, exhibited a diagnostic accuracy value of 72±3%, a sensitivity of 73±5%, and a specificity of 73±5%. An investigation of I-131 pharmacokinetics revealed that the maximum histogram intensity gradient (r=–0.48, p=0.08) and intensity entropy (r=–0.51, p=0.06) exhibited the highest Spearman correlation coefficient modulus with I-131 accumulation after 24 hours. CONCLUSIONS: The present study demonstrates the feasibility of using quantitative characterization of ultrasound images of nodal masses as a tool to monitor nodules before radioiodotherapy. This is with a view to subsequent adjunctive fine-needle aspiration puncture biopsy and prediction of I-131 accumulation after 24 hours.

Keywords

• endocrinology
• radioiodine therapy
• i-131
• nuclear medicine
• textural analysis
• radiomics
• machine learning
• ultrasound
• thyroid nodules