An interpretable MRI-based radiomics model predicting the prognosis of high-intensity focused ultrasound ablation of uterine fibroids

Chengwei Li; Zhimin He; Fajin Lv; Yang Liu; Yan Hu; Jian Zhang; Hui Liu; Si Ma; Zhibo Xiao

doi:10.1186/s13244-023-01445-2

An interpretable MRI-based radiomics model predicting the prognosis of high-intensity focused ultrasound ablation of uterine fibroids

Insights Imaging. 2023 Jul 19;14(1):129. doi: 10.1186/s13244-023-01445-2.

Authors

Chengwei Li¹, Zhimin He¹, Fajin Lv^{1

2}, Yang Liu¹, Yan Hu¹, Jian Zhang¹, Hui Liu¹, Si Ma¹, Zhibo Xiao^{3

4}

Affiliations

¹ State Key Laboratory of Ultrasound in Medicine and Engineering, College of Biomedical Engineering, Chongqing Medical University, Chongqing, China.
² Department of Radiology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, 400016, China.
³ State Key Laboratory of Ultrasound in Medicine and Engineering, College of Biomedical Engineering, Chongqing Medical University, Chongqing, China. 202530@cqmu.edu.cn.
⁴ Department of Radiology, The First Affiliated Hospital of Chongqing Medical University, Chongqing, 400016, China. 202530@cqmu.edu.cn.

Abstract

Background: Accurate preoperative assessment of the efficacy of high-intensity focused ultrasound (HIFU) ablation for uterine fibroids is essential for good treatment results. The aim of this study was to develop robust radiomics models for predicting the prognosis of HIFU-treated uterine fibroids and to explain the internal predictive process of the model using Shapley additive explanations (SHAP).

Methods: This retrospective study included 300 patients with uterine fibroids who received HIFU and were classified as having a favorable or unfavorable prognosis based on the postoperative nonperfusion volume ratio. Patients were divided into a training set (N = 240) and a test set (N = 60). The 1295 radiomics features were extracted from T2-weighted imaging (T2WI) and contrast-enhanced T1-weighted imaging (CE-T1WI) scans. After data preprocessing and feature filtering, radiomics models were constructed by extreme gradient boosting and light gradient boosting machine (LightGBM), and the optimal performance was obtained by Bayesian optimization. Finally, the SHAP approach was used to explain the internal prediction process.

Results: The models constructed using LightGBM had the best performance, and the AUCs of the T2WI and CE-T1WI models were 87.2 (95% CI = 87.1-87.5) and 84.8 (95% CI = 84.6-85.7), respectively. The use of SHAP technology can help physicians understand the impact of radiomic features on the predicted outcomes of the model from a global and individual perspective.

Conclusion: Multiparametric radiomic models have shown their robustness in predicting HIFU prognosis. Radiomic features can be a potential source of biomarkers to support preoperative assessment of HIFU treatment and improve the understanding of uterine fibroid heterogeneity.

Clinical relevance statement: An interpretable radiomics model can help clinicians to effectively predict the prognosis of HIFU treatment for uterine fibroids. The heterogeneity of fibroids can be characterized by various radiomics features and the application of SHAP can be used to visually explain the prediction process of radiomics models.

Keywords: HIFU; Machine learning; Magnetic resonance imaging; Radiomics; Uterine fibroid.