| Part1:Diagnostic value of placenta accrete spectrum disorders based on MRI radiomicsObjective The purpose of this study was to investigate the diagnostic value of Placenta accrete spectrum disorders based on MRI T2 WI radiomics by machine learning.Methods The imaging data of 168 patients with suspected placenta accretion who underwent MRI examination and later cesarean section were retrospectively analyzed.According to the postoperative results of cesarean section,MRI T2 WI images were used to extract the imaging features of the surface with placenta accretion and the surface without placenta accretion.The data were divided into training set(n=117)and test set(n=51)by stratified sampling in a ratio of 7∶3.Seven machine learning methods were used: Logistic Regression,Support Vector Machine,Random Forest,Stochastic Gradient Descent,Decision Tree7),K nearst Neighbors,and the Adaboost was generated by iterating the weak classifier for modeling classification diagnosis.The hyperparameters of the machine learning model were determined by the five-fold cross-validation,and the clinical model and the joint clinical-imaging model were constructed.Receiver operating characteristic(ROC)curve was used to evaluate the prediction efficiency of the model,and the area under the curve(AUC),accuracy,sensitivity and specificity were calculated.Each model is verified in the verification set.In addition,in addition to comparing the diagnostic effectiveness of different machine learning models with that of imaging diagnostic doctors,calibration curves were used to analyze the model efficacy,and decision curve analysis(DCA)was used to evaluate clinical practicability.Calibration curve and receiver operating characteristic(ROC)curve were used to analyze the efficacy of the model,and decision curve analysis(DCA)was used to evaluate its clinical practicability.Results Among the 168 patients included,76 had placenta accrete after cesarean section,and 92 had no placenta accrete.Based on 1688 comics features included in the image preprocessing,14 image comics features were selected for the construction of the model after LASSO,Selectbest and REF processing.The top 3 most important in the feature classification are: gray level co-occurrence matrix,first order statistics,gray level size zone matrix,gray level run length matrix.Seven kinds of classifier models were selected in the validation set(LR model AUC 09375,SVM model AUC0.9348,RF model AUC 0.8899,SGD model AUC0.9293,Adaboost model AUC0.9579,DT model AUC 0.8696,KNN model AUC)The diagnostic efficacy of placenta accrete was higher than that of imaging doctors(AUC=0.863).Calibration curve shows that the calibration degree of DT model is the best in verification set.When the threshold value of validation set DCA was 0 ～ 0.6,the clinical net benefit of RF,SVM,KNN,LR,SGD and DT models was greater than that of Adaboost model.Conclusion The machine learning model based on the radiomics features extracted from MRI sagittal T2 WI image segmentation of placenta and adjacent uterine wall can effectively identify potential placental tissue abnormalities and accurately diagnose PAS.The clinical-radiomics combined model established by the radiomics model combined with important clinical factors can further improve the diagnostic performance of the model.This method is significantly superior to the visual analysis of senior and junior diagnostic doctors,and is also a more differentiated objective assessment method of PAS than the current qualitative and subjective criteria.It can better assist imaging doctors to improve the diagnostic level of prenatal PAS,so as to provide a strong guarantee for the prevention of PAS risks.Part 2:Diagnostic value of MRI radiomics to predict the amount of bleeding during Caesarean section in pregnant womenObjective This study predicted the amount of bleeding during cesarean section by machine learning based on the omics features of MRI T2 WI images,providing a new idea and method for reducing the risk of bleeding during cesarean section clinically.Methods The imaging data of 168 patients who underwent MRI and caesarean section were retrospectively analyzed,including 76 patients with placenta accreta and 92 patients with normal placenta.The patients were divided into three groups according to the standard of emergency treatment of caesarean section bleeding.The first group: patients in the placenta accreta group were divided into 27 positive cases and 49 negative cases according to the bleeding warning line.The second group: the normal group was divided into 30 positive cases and 62 negative cases according to the early warning line of bleeding.The third group: patients in the placenta accreta group and the normal group were divided into 68 positive cases and 100 negative cases according to the bleeding warning line.The prediction model of intraoperative bleeding was constructed based on MRI T2 WI mapping placental data and five kinds of machine learning.The prediction efficiency of the model was analyzed by area under the curve,accuracy,sensitivity and specificity.In addition,combined with clinical characteristics(patient age,gestational age,weight,prenatal hemoglobin,placental thickness,cervical tube length),the normogram was jointly constructed.Results The prediction modeling analysis of cesarean section bleeding showed that machine learning predicted the amount of bleeding during cesarean section in the first group of patients: the LR model validation set AUC was 0.3472,the sensitivity was 41.76%,the specificity was 33.33% and the accuracy was 33.33%.The validation set AUC of RF model was 0.7639,with sensitivity 83.33%,specificity 16.67% and accuracy 33.33%.The validation set AUC of SGD model was 0.375,with sensitivity 41.67%,specificity 16.67% and accuracy 33.33%.The validation set of KNN model had an AUC of 0.6667,sensitivity of 83.33%,specificity of 16.67% and accuracy of 61.11%.Prediction of blood loss during cesarean section in the second group: LR model validation set AUC was 0.7583,sensitivity was 3.75%,specificity was 86.67%,accuracy was 69.57%;The AUC of SVM model validation set was 0.575,the sensitivity was 6.25%,the specificity was 66.67% and the accuracy was 65.22%.The validation set of RF model had an AUC of 0.7917,a sensitivity of 7.5%,a specificity of 73.33% and an accuracy of 73.91%.The AUC of SGD model validation set was 0.7583,the sensitivity was 3.75%,the specificity was 86.67%,and the accuracy was 69.57%.The validation set AUC of KNN model was 0.7,with sensitivity 0,specificity 100% and accuracy 65.22%.Prediction of blood loss during cesarean section in the third group: LR model validation set AUC was 0.7381,sensitivity was 76.19%,specificity was 55%,accuracy was 65.85%;The AUC of SVM model validation set was 0.7571,the sensitivity was 57.14%,the specificity was 70% and the accuracy was 63.41%.The AUC of RF model validation set was 0.7905,the sensitivity was 76.19%,the specificity was 60% and the accuracy was 68.29%.The AUC of SGD model validation set was 0.7167,with sensitivity of 61.9%,specificity of 55% and accuracy of 58.54%.The validation set of KNN model had an AUC of 0.7131,sensitivity of 71.43%,specificity of 60% and accuracy of 65.85%.Conclusion Machine learning is not effective in predicting the amount of blood loss during cesarean section in patients with placenta accreta group,but the prediction of the amount of blood loss during cesarean section in patients with combined group is the highest.Therefore,MRI texture analysis and machine learning can help radiologists to accurately predict intraoperative bleeding in patients with cesarean section.It facilitates the preparation of blood products and the provision of relevant hemostatic measures.In addition,this study showed that placenta accreta was not the main cause of postpartum massive bleeding as long as the patients were accurately diagnosed before surgery. |
PDF Full Text Request