Establishment of Noninvasive Prediction Models for the Diagnosis of Uterine Leiomyoma Subtypes

Abstract

OBJECTIVE: To establish prediction models for the diagnosis of the subtypes of uterine leiomyomas by machine learning using magnetic resonance imaging (MRI) data. METHODS: This is a prospective observational study. Ninety uterine leiomyoma samples were obtained from 51 patients who underwent surgery for uterine leiomyomas. Seventy-one samples (49 mediator complex subunit 12 [MED12] mutation–positive and 22 MED12 mutation–negative leiomyomas) were assigned to the primary data set to establish prediction models. Nineteen samples (13 MED12 mutation–positive and 6 MED12 mutation-negative leiomyomas) were assigned to the unknown testing data set to validate the prediction model utility. The tumor signal intensity was quantified by seven MRI sequences (T2-weighted imaging, apparent diffusion coefficient, magnetic resonance elastography, T1 mapping, magnetization transfer contrast, T2* blood oxygenation level dependent, and arterial spin labeling) that can estimate the collagen and water contents of uterine leiomyomas. After surgery, the MED12 mutations were genotyped. These results were used to establish prediction models based on machine learning by applying support vector classification and logistic regression for the diagnosis of uterine leiomyoma subtypes. The performance of the prediction models was evaluated by cross-validation within the primary data set and then finally evaluated by external validation using the unknown testing data set. RESULTS: The signal intensities of five MRI sequences (T2-weighted imaging, apparent diffusion coefficient, T1 mapping, magnetization transfer contrast, and T2* blood oxygenation level dependent) differed significantly between the subtypes. In cross-validation within the primary data set, both machine learning models (support vector classification and logistic regression) based on the five MRI sequences were highly predictive of the subtypes (area under the curve [AUC] 0.974 and 0.988, respectively). External validation with the unknown testing data set confirmed that both models were able to predict the subtypes for all samples (AUC 1.000, 100.0% accuracy). Our prediction models with T2-weighted imaging alone also showed high accuracy to discriminate the uterine leiomyoma subtypes. CONCLUSION: We established noninvasive prediction models for the diagnosis of the subtypes of uterine leiomyomas by machine learning using MRI data.