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

Area of Science:

• Nuclear Medicine
• Radiology
• Oncology

Background:

• Thyroid cancer diagnoses are increasing, with many detected incidentally (thyroid incidentaloma, TI) via imaging like PET/CT.
• Conventional parameters like SUVmax are insufficient for characterizing TI.
• Radiomics, analyzing image texture, offers a quantitative method to identify subtle features.

Purpose of the Study:

• To evaluate the effectiveness of radiomics features and an AI-based algorithm in detecting malignancy in [18F]FDG-avid thyroid incidentalomas.
• To compare the diagnostic performance of radiomics with conventional SUVmax measurements.

Main Methods:

• Analysis of 50 patients with [18F]FDG-avid thyroid incidentalomas who underwent fine needle aspiration biopsy.
• Application of an XGBoost model utilizing radiomics features extracted from PET/CT images.
• Comparison of the AI-radiomics model's performance against SUVmax values.

Main Results:

• Out of 50 patients, 11 (22.0%) [18F]FDG-avid nodules were malignant.
• The XGBoost radiomics model achieved a performance of 0.846 (95% CI: 0.737-0.956).
• Performance of the radiomics model was similar to SUVmax (0.797; 95% CI: 0.622-0.973), with no statistically significant difference (p=0.60).

Conclusions:

• An AI-based algorithm using radiomics features can potentially detect thyroid nodule malignancy.
• No statistically significant difference was found between the AI-radiomics approach and the conventional SUVmax measurement.
• Radiomics offers a complementary tool for characterizing thyroid incidentalomas, but further validation is required.