Abstract
BACKGROUND: Thyroid diseases are the second most common hormonal disorders, necessitating accurate diagnostics. Advances in artificial intelligence and radiomics have enhanced diagnostic precision by analyzing quantitative imaging features. However, reproducibility challenges arising from factors such as the field-of-view (FOV) zooming and segmentation variability limit the clinical application of radiomic-based models. AIM: This study focuses on evaluating the impact of segmentation and FOV zooming on the reproducibility of radiomic features and improved performance of machine learning (ML) when using reproducible features for classification of thyroid scintigraphy images into normal, diffuse goiter (DG), multinodular goiter (MNG), and thyroiditis. PATIENTS AND METHODS: A retrospective analysis was conducted on 872 thyroid scintigraphy cases from 3 centers. Radiomic feature reproducibility was assessed using the intraclass correlation coefficient (ICC), with robust features (ICC≥0.80) identified under segmentation and zooming conditions. Four ML training scenarios were implemented to train models on Center A data, including (1) all, (2) zoom-robust, (3) segmentation-robust, and (4) mutually robust features, with 3 feature selection methods and 7 classifiers. Models were validated on external data sets (centers B and C). RESULTS: FOV zooming significantly reduced feature reproducibility (ICC≥0.80: 49%), while segmentation effects were minimal (ICC≥0.80: 96%). Models trained on mutually robust features outperformed those trained using all features. Boruta-MLP achieved the highest accuracy (0.71, P -value <0.001 vs. all features) in zoomed data sets, and RFE-MLP performed best (0.69, P -value <0.001 vs. all features) in the baseline data set, with Gray-Level Co-occurrence Matrix (GLCM) features frequently selected. CONCLUSIONS: Utilizing robust radiomic features significantly improved the performance of ML models in thyroid disease classification, enabling more accurate and generalizable diagnostic outcomes across diverse data sets.