Multi-modal feature integration for thyroid nodule prediction: Combining clinical data with ultrasound-based deep features

This study presents an advanced machine learning (ML)-based framework for accurate risk stratification of thyroid nodules by integrating clinical, radiological, and deep imaging features. We analyzed data from 580 patients with thyroid nodules, categorized from TIRADS 2 to 5. Clinical and radiological features (e.g., nodule size, TIRADS category, echogenicity) were combined with deep imaging features extracted from ultrasound scans using EfficientNet-B0. Predictive models were developed using XGBoost, Random Forest, and Support Vector Machine (SVM), resulting in clinical-only, imaging-only, and hybrid models. Additionally, a stacking-based meta-model integrated predictions from all three models to enhance performance. Model evaluation metrics included accuracy, F1-score, and AUC-ROC, with hyperparameter tuning applied to optimize outcomes. The clinical models showed strong predictive ability, with XGBoost achieving 81% accuracy, 0.85 AUC-ROC, and an F1-score of 0.82. Imaging models demonstrated the value of deep feature extraction, reaching up to 79% accuracy and 0.83 AUC-ROC. The hybrid model improved predictions further, with XGBoost achieving 85% accuracy and 0.87 AUC-ROC. The stacking ensemble model provided the best performance, achieving 87% accuracy, an F1-score of 0.87, and 0.90 AUC-ROC, demonstrating the benefits of multi-modal feature integration. Our study shows that combining clinical, radiological, and deep imaging features significantly enhances the prediction of thyroid nodule malignancy. The stacking-based framework offers a scalable, reproducible tool to support more accurate clinical decision-making, reduce unnecessary biopsies, and improve diagnostic precision.