AI may help to predict thyroid nodule malignancy based on radiomics features from [18F]FDG PET/CT.

Background: The number of thyroid cancer diagnoses has been increasing for several decades, with a significant part of cases being detected incidentally (thyroid incidentaloma, TI) by imaging studies performed for reasons other than thyroid disease, including PET/CT with [¹⁸F]FDG. The chacteristics of the detected TI cannot be determined solely on the basis of conventional parameters used in everyday clinical practice, such as SUV_max. In recent years, there has been a growing interest in radiomics, which is a quantitative method of analyzing radiological images based on the analysis of image texture. Textural analysis may be helpful, as it allows to characterize features invisible to the physician with the naked eye.

Results: Of the 54 patients who presented focal [¹⁸F]FDG-avid TI and had subsequent fine needle aspiration biopsy, 4 patients were excluded from the analysis due to the unavailability of the final diagnostic information. Hence, in the final analysis, data from 50 patients were used (39 females and 11 males) with a mean age of 58.5 ± 11.26. Of these 50 patients, 11 (22.0%) [¹⁸F]FDG-avid nodules were diagnosed as malignant. The performance of the XGBoost model in assessing [¹⁸F]FDG-avid TI was similar (0.846 [confidence interval, CI, 95% 0.737-0.956]) to SUV_max (0.797 [CI 95%: 0.622-0.973]; p = 0.60).

Conclusions: With an AI-based algorithm using radiomics features it is possible to detect the malignancy of thyroid nodule. However, no statistically significant differences were observed between the AI and radiomics approach, and when using a conventional measure, i.e., SUV_max.