Abstract
The intratumoral and peritumoral architectural heterogeneities of papillary thyroid carcinoma (PTC) are important in preoperative prediction of gross extrathyroidal extension (Gross ETE). This study systematically evaluated and compared the predictive efficacies of deep learning, radiomics, and their combined approach (Deep Learning-Radiomics, DLR) in predicting Gross ETE in PTC patients using ultrasound imaging. This retrospective study from three hospitals, between 01/01/2018, and 12/31/2022, included 4,542 PTC patients, divided into training (n = 3,179) and testing (n = 1,363) sets in a 7:3 ratio. Preoperative ultrasound images and clinical data were collected to develop radiomics and deep learning models based on different tumor expansion regions (5/10/15/20 pixels). A nomogram prediction model was developed by integrating multi-regional radiomics features and key clinical parameters. Model performance was assessed using metrics such as the area under the curve (AUC), sensitivity, and specificity. Feature importance was evaluated using SHapley Additive exPlanations (SHAP) analysis, and model interpretability was analyzed with Gradient-weighted Class Activation Mapping (Grad-CAM). In the test cohort, the radiomics model with 15 pixel expansion (AUC: 0.796) and the ResNet101 deep learning model (AUC: 0.832) showed optimal performance. The DLR model incorporating 15 pixel peritumoral features (DLRexpand15) combined with clinical parameters achieved superior predictive performance (AUC: 0.849, accuracy: 0.857, and specificity: 0.888). SHAP analysis identified deep learning features as the primary predictors in the fusion model, while Grad-CAM visualization confirmed spatial concordances between model-activated regions and histopathological invasion patterns. The DLRexpand15-based nomogram integrating clinical indicators provided an effective tool for preoperative prediction of Gross ETE in PTC patients. Strategic incorporation of peritumoral information significantly enhanced the predictive capacity of both radiomics and deep learning models. This multimodal approach provided clinically useful insights for surgical planning and risk stratification.