Machine learning-based ultrasound radiomics for prediction of 6-month local thermal ablation response in benign thyroid nodules: a multicenter study.
Authors
Affiliations (5)
Affiliations (5)
- Department of Medical Ultrasound, Center of Minimally Invasive Treatment for Tumor, Shanghai Tenth People's Hospital, Ultrasound Research and Education Institute, Clinical Research Center for Interventional Medicine, School of Medicine, Tongji University, Shanghai, China.
- Shanghai Engineering Research Center of Ultrasound Diagnosis and Treatment, Shanghai, China.
- Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University, Shanghai, China.
- Department of Ultrasound, Zhongshan Hospital, Institute of Ultrasound in Medicine and Engineering, Fudan University, Shanghai, China.
- Department of Ultrasound, Sichuan Provincial People's Hospital, University of Electronic Science and Technology of China, Chengdu, China.
Abstract
This study aimed to develop a machine learning (ML)-based ultrasound (US) radiomics model for prediction of 6-month local treatment response (LTR) ofthermal ablation (TA) for benign thyroid nodules (BTNs). Between January 2018 and July 2021, a total of 388 patients who underwent US-guided TA in three centers were included. US radiomics features were extracted from preoperative grayscale US images and data dimensionality reduced using principal component analysis and least absolute shrinkage and selection operator. Then, support vector machine (SVM), logistic regression, a decision tree, K-nearest neighbors, and random forest were applied to selected key US radiomics features for distinguishing a volume reduction ratio (VRR) ≥50% or <50%. Factors affecting 6-month LTR were assessed using multivariate logistic regression to construct a clinical model. Receiver operating characteristic curves were plotted to compare the predictive performance between radiomics-based ML and clinical models. At 6 months post-ablation, patients with VRR ≥ 50% were 75.8% (292/372) in the training and internal test cohorts and 59.4% (19/32) in the external test cohort. Finally, 10 US radiomics features were selected for analysis. Solidity was the only independent clinical predictor associated with VRR<50%. Among the five algorithms, the SVM model achieved the optimal predictive efficacy, with an area under the curve (AUC) = 0.81 in the internal test set. The AUC of the SVM-based US radiomics model was significantly higher than that of the clinical model in both internal (0.81 vs. 0.63, P< 0.05) and external test cohorts (0.77 vs. 0.54, P< 0.05). The SVM-based US radiomicsmodel yielded a satisfactory performance for predicting the 6-month LTR, outperforming the clinical model, and facilitating decision-making in favor of TA.