Super-resolution ultrasound radiomics for pre-FNA prediction of nondiagnostic (Bethesda I) thyroid nodules.

Introduction: We evaluated whether a generative adversarial network (GAN)-based super-resolution (SR) radiomics pipeline could improve the pre-fine-needle aspiration (FNA) prediction of nondiagnostic (Bethesda I) outcomes in thyroid nodules compared with normal-resolution (NR) ultrasound analysis. Methods: This single-center retrospective study included 437 patients with thyroid nodules, with 338 assigned to the development cohort and 99 to a temporally independent validation cohort. Two-dimensional B-mode ultrasound images were processed using a pretrained GAN-based SR model in inference-only mode. Handcrafted radiomic features were extracted using PyRadiomics and subjected to a multistep selection pipeline comprising the Mann-Whitney U test, Spearman correlation-based redundancy pruning, random forest-based recursive feature elimination (RFE), and least absolute shrinkage and selection operator (LASSO) regression. Four machine-learning classifiers, including random forest (RF), LightGBM, ExtraTrees, and XGBoost, were trained using five repeated stratified 80/20 splits. Post hoc probability calibration was performed using Platt (sigmoid) scaling fitted on the training set and applied unchanged to the independent validation set. In addition, qualitative error analysis was conducted to characterize false-positive and false-negative cases and to assess the potential effect of operator experience on model performance. Results: The SR-based models consistently outperformed the NR-based models. The SR-RF model achieved an area under the receiver operating characteristic curve (AUC) of 0.808 in training and 0.733 in internal testing, compared with 0.672 and 0.596, respectively, for the NR-RF model. In the independent validation cohort, the SR-RF model yielded an AUC of 0.7435. Platt calibration improved the Brier score from 0.1885 to 0.1702 without affecting AUC, indicating improved probability reliability. Across classifiers, RF and LightGBM showed the most balanced overall performance. Qualitative error analysis showed that false-positive cases were often associated with cystic or heterogeneous echotexture and posterior acoustic artifacts, whereas false-negative cases involved isoechoic solid nodules with subtle margins and limited region-of-interest (ROI) coverage. No systematic effect of operator experience on model performance was observed. Discussion: GAN-based SR radiomics significantly improves the pre-FNA prediction of nondiagnostic cytology in thyroid nodules. When combined with post hoc calibration, this approach provides more reliable individualized risk estimates and may help reduce unnecessary repeat procedures.