Abstract:

Objective To observe the value of multi-sequence magnetic resonance imaging (MRI) radiomics in predicting the efficacy of concurrent chemoradiotherapy (CCRT) in locally advanced cervical squamous cell carcinoma (CSCC) patients. Methods Clinical data of 100 CSCC patients underwent CCRT treatment were selected. In order to better validate the performance of the model, patients were randomly divided into the training set (70 cases) and the validation set (30 cases) in a 7∶3 ratio. According to the efficacy criteria for solid tumors, patients were divided into the complete response (CR) group (n=16) and the partial response (PR) group (n=14). Examination images of cross-sectional DWI, T₂WI and enhanced T₁WI were collected from all patients before treatment. ITK-SNAP software package combined with three sequences were used to outline ROI, and the open source software PyRadiomics was used to extract image omics features. For MRI omics features, the minimum redundancy maximum correlation (mRMR) algorithm was used to analyze and screen out the first 30 main features, and then the minimum absolute contraction and selection method (Lasso) based on 10-fold cross-validation was used to reduce dimensionality to screen the non-zero coefficient features. According to the weighting coefficient of Lasso-Logistic regression model in the training set, patient omics labels were calculated. Logistic regression analysis was used to construct a prediction model based on DWI, T₂WI and T₁WI sequence prediction models and multiple sequenomics labels. Receiver operating characteristic (ROC) curves evaluated the predictive value of each omics model for CCRT treatment in patients with locally advanced CSCC. Results There were 38 cases in the CR group and 32 cases in the PR group in the training set. There were 16 cases in the CR group and 14 cases in the PR group in the validation set. There were no significant differences in patient age, FIGO stage, differentiation degree, maximum lesion diameter and menstrual status between the CR group and the PR group in the training and validation sets. A total of 851 imaging features were extracted from the ROI target area. After the first 30 features were retained by mRMR algorithm, 3 CR-related features were selected from the 851 imaging omics features of each individual sequence by Lasso algorithm and 10-fold cross-validation. Eight CR related features were selected from 2 553 features after the combination of the three sequences. ROC curve results showed that in the training set and validation set, the AUC of multiple sequences combined to predict the therapeutic effect of CCRT in patients with locally advanced CSCC was 0.971 and 0.946, respectively, which was higher than that of T₁WI, T₂WI and DWI single sequence prediction (training set Z=2.683, 2.046, 2.817, P<0.05; verification set Z=2.075, 2.117, 2.005, P<0.05). Conclusion The multi sequence MRI radiomics model has high predictive value for the efficacy of CCRT treatment in locally advanced CSCC patients.

Key words: uterine cervical neoplasms, carcinoma, squamous cell, magnetic resonance imaging, treatment outcome, radiomics