Multiparametric magnetic resonance imaging (MRI)-based interpretable radiomics machine learning model for outcome prediction in adenomyosis treated by high-intensity focussed ultrasound

AIM: This study sought to design a multiparametric magnetic resonance imaging (MRI) radiomics machine learning model capable of predicting high-intensity focussed ultrasound (HIFU) ablation outcomes in adenomyosis, with the further implementation of the Shapley Additive Explanation (SHAP) approach to explain the internal predictive processes of the machine learning model. MATERIALS AND METHODS: Data from 170 adenomyosis patients treated via HIFU were analysed retrospectively, randomising these patients into training and testing cohorts in a 7:3 ratio. T2-weighted images (T2WIs) and apparent diffusion coefficient (ADC) images were used to manually outline the volume of interest. Four machine learning algorithms including decision tree (DT), logistic regression (LR), random forest (RF), and support vector machine (SVM) algorithms were built based on ADC (ADC model), T2WI (T2 model), and ADC+T2WI (dual-sequence magnetic resonance imaging [dsMRI] model). Model performance was compared with receiver operating characteristic (ROC) curves and the DeLong test, after which the interpretation and visualisation of the best-performing model were achieved using an SHAP. RESULTS: In the testing cohort, the dsMRI model based on the RF algorithm outperformed the other models, with respective area under the curve (AUC) values of 0.902 and 0.823 in the training and testing cohorts, respectively. For the ADC model, RF achieved the highest AUC of 0.701 in the testing cohorts. For the T2 model, RF achieved the highest AUC of 0.747 in the testing cohorts. The internal radiomics feature-based predictive processes of the dsMRI model were successfully explained at the global and local levels with the SHAP method. CONCLUSION: The radiomics machine learning-based dsMRI model developed herein can effectively predict HIFU ablation outcomes in adenomyosis.