Value of machine learning model based on MRI radiomics in predicting histological grade of cervical squamous cell carcinoma

Abstract

Objective To explore the predictive value of different machine learning models based on MRI radiomics combined with clinical features for histological grade of cervical squamous cell carcinoma. Methods Clinical data of 150 patients with cervical squamous cell carcinoma confirmed by pathological biopsy were retrospectively analyzed. They were randomly divided into the training set and validation set at a ratio of 4∶1. Features were extracted from the regions of interest of T2WI fat suppression sequence （FS-T2WI） and enhanced T1WI （delayed phase）. After dimensionality reduction and feature selection， logistic regression （LR）， support vector machine （SVM）， naïve Bayes （NB）， random forest （RF）， Light Gradient Boosting Machine （LightGBM）， K-nearest neighbor （KNN） were used to construct a radiomics model for predicting the histological grade of cervical squamous cell carcinoma. The area under the receiver operating characteristic （ROC） curve （AUC） was used to evaluate the predictive performance of the six models. Univariate and multivariate logistic regression analyses were performed to predict the independent risk factors， and a combined model of clinical and radiomics was established. The differences of each model were compared by AUC， and the clinical value of the model was evaluated by decision curve （DCA）. Results In the radiomics model， the LightGBM model had the largest AUC （0.910 in the training set， and 0.839 in the validation set）. The AUC of clinical features combined with LightGBM model was the largest （0.935 in the training set， and 0.888 in the validation set）， which was higher than those of clinical model （0.762 in the training set， and 0.710 in the validation set） and LightGBM radiomics model. Conclusions The LightGBM model has a high predictive value in the radiomics model. The combined model has the optimal DCA effect and the highest clinical net benefit. The combined prediction model combining radiomics and clinical features has good predictive value for cervical squamous cell carcinoma with low differentiation， providing a non-invasive and efficient method for clinical decision-making.

Keywords

• cervical cancer|radiomics|histological grade|magnetic resonance imaging|machine learning