Automated multi-class classification of thyroid nodules in ultrasound imaging using transformer-based segmentation and hybrid feature learning

To develop and validate an end-to-end machine learning pipeline for automated multi-class classification of thyroid nodules in ultrasound imaging, using transformer-based segmentation and hybrid radiomic-deep feature integration to enhance clinical accuracy and reproducibility. In this multi-center study, 2654 ultrasound cases from five hospitals were used for model development, and 873 additional cases from an independent center were used for external validation. Thyroid nodules were segmented using four architectures: UNETR, nnU-Net, Swin-UNet, and UNet. From the segmented regions, handcrafted radiomic features and deep features extracted via Vision Transformer encoder layers were obtained. Features were filtered using ICC ≥0.75, followed by variance and correlation-based refinement. Three feature selection methods—Lasso, PCA, and Mutual Information—were evaluated. Classification was performed using XGBoost, Random Forest, and TabTransformer across six TI-RADS categories. Five-fold stratified cross-validation and external testing ensured robustness. Segmentation was assessed using Dice, Jaccard, and Hausdorff metrics; classification performance was evaluated via accuracy, AUC, and recall. UNETR achieved the highest segmentation performance and enabled the most accurate classification. The best outcome was observed with radiomic features selected via Lasso and classified with XGBoost (external accuracy: 93.0 %, AUC: 93.6 %, recall: 92.0 %). Deep features showed comparable results (accuracy: 92.8 %). Q-value analysis confirmed statistical superiority of the best-performing models. Differences across segmentation models significantly impacted classification performance, highlighting the importance of boundary quality. All models demonstrated strong generalizability and minimal overfitting. The study demonstrates the feasibility and clinical value of a fully automated pipeline for TI-RADS-based thyroid nodule classification. The proposed framework is generalizable, interpretable, and suitable for integration into real-time diagnostic systems.