A Study Of Predicting Pituitary Adenoma Consistency Based On Mri Radiomics

Objective: To investigate the clinical application value of noninvasive preoperative prediction of pituitary adenoma consistency based on MR imaging and radiomics.The prediction efficiency of each radiomics model for pituitary adenoma consistency was compared,so as to obtain the best prediction model and improve the accuracy of preoperative prediction of pituitary adenoma consistency.Materials and Methods: The clinical and imaging data of 92 patients with pituitary adenoma confirmed by surgery and pathology in our hospital from June2018 to November 2020 were analyzed retrospectively.According to the intraoperative conditions,all the pituitary adenomas were divided into 61 cases of soft group and 31 cases of firm group according to the consistency classification standard of Mahmoud.According to the partitioning principle of small data sets,patients were divided into training group(n = 74)and test group(n = 18)by the method of stratified random sampling in a ratio of 8:2.Clinical data of all patients were collated and 3.0T MR magnetic resonance data were derived,and then the radiomics characteristics of tumor volume were extracted based on T2 WI using3D-Slicer software,a texture analysis tool.Recursive feature elimination(RFE)and Pearson correlation coefficient(PCC)were used for selecting features.Multi-layer Perceptron(MLP),Support Vector Machine(SVM),Logical Regression(LR)and lasso constrained Logical Regression(LR-lasso)classifiers were used to construct the radiomic models.Five-fold cross-validation was used to train and validate the performance of the radiomics model.We used Receiver operating characteristic(ROC)curve to evaluate the predictive efficacy of different radiomic models for pituitary adenoma consistency,The area under the curve(AUC),accuracy(ACC),specificity(Spec),sensitivity(Sens),negative predictive value(NPV),positive predictive value(PPV)of each radiomics model were calculated,and De Long test was used to compare the difference of area under the curve(AUC).Results:(1)There was no significant difference between the soft and firm groups in gender,age,tumor shape and tumor maximum diameter;(2)AUC of SVM training set and test set were 0.811 and 0.819,respectively.AUC of training set and test set of LR model were 0.806 and 0.847,respectively.AUC of training set and test set of LR-lasso model were 0.809 and 0.833,respectively.AUC of MLP model training set and test set were 0.800 and 0.778,respectively.(3)Pair comparison of the ROC curves of the four models showed that there was no statistical significance in the AUC of the four models(all P values were greater than 0.05).Conclusions:(1)SVM,LR,LR-lasso and MLP radiomics model based on magnetic resonance T2 WI sequences are helpful for noninvasive preoperative prediction of pituitary adenoma consistency;(2)The four radiomics model were equally effective in predicting the tumor consistency of pituitary adenoma.