CT-Based Explainable Machine Learning for Predicting Benign and Malignant Thyroid Nodules: A Multi-center Study

海军 何^1,2Mingquan Luo^1,2Kai Hu^1,2Tengfei Ke³Juntao Yang⁴Xinyue Zi⁵Tingting Jiang⁶Liangping Yong²Tong Chen²Jun Chen²Zhengliang He²Qiangrong Gao²Zhoubin Liang²Yang Jing^7*Bin Yang⁸

• ¹Nanbu County People's Hospital, 南充市, China
• ²Radiology Department, People's Hospital of Nanbu County, Nanchong City, Sichuan Province, China
• ³Yunnan Cancer Hospital，The Third Affiliated Hospital of Kunming Medical University，Peking University Cancer Hospital Yunnan Campus,China, Kunming, China
• ⁴Radiology Department, Dali Bai Autonomous Prefecture People's Hospital, Dali City, Yunnan Province,China, Dali, China
• ⁵The First Affiliated Hospital of Dali University, Dali, China
• ⁶lood Transfusion Department, People's Hospital of Nanbu County, Nanchong City, Sichuan Province, Nanchong, China
• ⁷Huiying Medical Technology Co., Ltd, Beijing, China
• ⁸Medical Imaging Centre，Kunming First People's Hospital Medical Imaging Centre, Kunming, China

Objective: To construct a CT-based explainable machine learning model for pre operative prediction of thyroid nodule benignity or malignancy, aiming to provi de a more accurate tool for clinical decision-making and management. Materials and Methods: A retrospective study included 370 patients with thyroi d nodules confirmed by pathology from three centres, divided into a training s et (n=229) and an internal validation set (n=100) in a 7:3 ratio, with patients f rom the third centre serving as an external validation set (n=41). Radiomics fe atures were extracted from preoperative CT images, and the optimal features w ere selected to construct a radiomics score (Rad_Score). Clinical risk factors w ere identified using univariate and multivariate logistic regression. LR and SV M algorithms were used to establish three models: a clinical model, an imagin g model, and a combined model (integrating clinical factors and Rad_Score). T he combined model was visualized using SHAP(SHapley Additive exPlanations) analysis. Model performance was evaluated using receiver operating characteristi c (ROC) curves, calibration curves, and decision curve analysis (DCA). Results: Seventeen features were ultimately selected for Rad_Score calculation. The combined model demonstrated the best performance, with the LR combine d model achieving AUC values of 0.962, 0.913, and 0.914 in the training set, internal validation set, and external validation set, respectively, all higher than t he LR clinical model and LR radiomics model; and the LR combined model o utperforms the SVM combined model (0.953, 0.885, 0.842). SHAP analysis rev ealed the relative importance of the key feature (Rad_score) in model predictio n, enhancing model transparency. Conclusion: The combined model performs better under the LR algorithm. Com bined with SHAP explainabel analysis, it provides a non-invasive, efficient, and transparent tool for preoperative differentiation of benign and malignant thyroi d nodules, potentially optimising individualized clinical management.