Prediction Of Lymphovangivascular Invasion Of Cervical Cancer Based On Multisequence MRI Radiomics

Background: Cervical cancer is one of the common cancers in women and has been gradually increasing in recent years.Early detection and accurate evaluation can enable clinical development of more reasonable plans for effective treatment of cervical cancer,thereby improving the quality of life and survival rate of patients.At present,the pretreatment evaluation of cervical cancer mainly uses conventional imaging examinations such as ultrasound,CT,or MRI to stage the tumor,and uses hysteroscopy to determine the histological type of the tumor.However,the accuracy of comprehensive evaluation is relatively limited,especially in the evaluation of micro lesions.Radiomics is a technology that extracts lesion feature information from images,with the advantage of intuitively and quantitatively displaying deep level features of lesions.It has the potential to non-invasive predict tumor staging,histological type,radiotherapy and chemotherapy efficacy,lymph node status,and tumor recurrence.Objective: To construct a clinical imaging omics model by analyzing patient clinical data and preoperative multi parameter magnetic resonance imaging features,and non-invasive predict lymphatic space infiltration(LVSI)in cervical cancer,in order to assist in clinical diagnosis and treatment.Research Method: This study retrospectively collected imaging images and clinical data of stage Ia-IIIc cervical cancer,and re sampled patients.Sketch each tumor area of interest on magnetic resonance imaging data,including axial T1 WI,axial T2WI-FS,DWI,and ADC images.Extract features based on four different sequences of lesion regions of interest.Divide the overall data into training and testing groups in a ratio of 7:3.Secondly,the extracted features are screened and dimensionally reduced in the training group.Firstly,perform variance analysis to preserve imaging features with high variance.Secondly,we used correlation analysis to compare the correlation between LVSI negative and LVSI positive features,and removed features with coefficients greater than 0.7 to further remove redundancy from the data.Then,the Mann Whitney U test was used to perform a univariate difference test on all features in the training group,and features with statistical significance(P<0.05)were selected between LVSI negative and LVSI positive groups.Using the Least absolute shrinkage and selection operator(LASSO),further dimensionality reduction of imaging omics features is performed,important features for predicting LVSI state are selected,and the Rad-score of each patient is obtained.A single sequence model and a joint model are established.Perform single factor and multivariate logistic regression analysis on clinical characteristic variables to obtain the most relevant candidate characteristic variables.Establish a combined MRI imaging omics model by combining single sequence model and joint sequence model Rad-score with preoperative clinical candidate feature variables.For the evaluation of the predictive performance of each model,AUC value,specificity,and sensitivity are used as predictive indicators.Based on the optimal model,establish a column chart to visualize the results for clinical use.The calibration curve and decision analysis curve were used as evaluation indexes to evaluate the fitting degree and clinical application value of the model.Result: This study established a total of 4 single sequence MRI sequence models,1 joint sequence MRI sequence model,and a combined model of Rad-score combined with clinical features for 5 MRI sequences.Firstly,compare the predictive performance of T1 WI,T2WI-FS,DWI,ADC,and joint sequence models.In the training group,the AUC value of the T1 WI sequence model was 0.862(0.762-0.961);The AUC value of the T2WI-FS sequence model is 0.838(0.721-0.954);The AUC value of the DWI sequence model is 0.866(0.768-0.964);The AUC value of the ADC sequence model is 0.833(0.720-0.947);The AUC value of the joint sequence model is 0.860(0.757-0.963).The DWI sequence model has the best performance among them.The De Long test was performed on each model,and there was no statistically significant difference in the predictive effect of LVSI state.We used univariate and multivariate regression for clinical characteristic variables,and reserved HGB,SCC Ag,and family history of tumor as independent clinical predictors to establish a clinical model.We selected single sequence and combined sequence Rad-scores with clinical feature variables to construct 5 composite models.Model 1(T1WI sequence model+clinical model)had an AUC value of 0.933(0.862 0.961);The AUC value of Model 2(T2WI sequence model+clinical model)is 0.905(0.823-0.986);The AUC value of Model 3(DWI sequence model+clinical model)is 0.939(0.881-0.996);The AUC value of Model 4(ADC sequence model+clinical model)is 0.936(0.876-0.996);The AUC value of Model 5(Joint Sequence Model+Clinical Model)is 0.927(0.860-0.994).Among them,Model 3 has the highest AUC value.The De Long test results showed no significant differences between the combination models.The DWI sequence model has the highest AUC value in both single and joint sequences,while Model 3 has the highest AUC value in the combined model.Moreover,Model 3 composed of the DWI sequence model and clinical model is superior to the single sequence model and other combined models.Therefore,we ultimately chose Model 3 composed of the DWI sequence model and clinical model as the prediction model for LVSI state.The AUC values of Model 3 in the training and testing groups were 0.939(0.881 0.996)and 0.742(0.531 0.953),respectively.We conducted Delong tests on Model 3,Clinical Model,and DWI Sequence Model,and the results showed significant differences(P=0.010 and 0.046)between the Clinical Model and DWI Sequence Model,indicating that Model 3 has a good predictive effect on LVSI in cervical cancer.Draw a visual column chart for clinical use.The calibration curve based on Model 3 shows satisfactory consistency between the prediction and observation probabilities of LVSI state in both the training and testing groups.The decision curve indicates that combining models within most threshold probability ranges will bring more clinical benefits.Conclusion: The combination model of Rad-Score of DWI sequence and clinical characteristic variables has good predictive effect on LVSI of cervical cancer.The plotted column chart can conveniently and personalized predict LVSI status of cervical cancer patients,facilitating clinical and patient treatment decisions.