Application Of Machine Learning Based On Multiparametric MRI In The Differential Diagnosis Of Uterine Sarcoma And Atypical Leiomyoma

Objective:To explore the feasibility and effectiveness of machine learning(ML)based on multiparametric magnetic resonance image(mp-MRI)features extracted transfer learning combined with clinical parameters to differentiate uterine sarcoma from atypical leiomyoma(ALM).Methods: Clinical and imaging data were retrospectively collected from 172 patients with pathologically confirmed uterine tumors,including86 cases of uterine sarcoma and 86 cases of atypical leiomyoma(ALM).All patients underwent preoperative mp-MRI examination,including T2-weighted imaging(T2WI)and diffusion-weighted imaging(DWI).Four pre-trained convolutional neural networks(CNNs),namely Inception V3,Resnet50,Inceptionresnetv2 and Xception,were used to extract deep learning features from the delineated region of interest(ROI),and Pyradiomics was used to extract radiomics features from delineated ROI.The performances of two feature extraction methods,transfer learning and radiomics,were compared.Feature selection was performed using the least absolute shrinkage selector operator(LASSO).The training and test sets were randomly divided in a ratio of 7:3,random forest(RF)was adopted as the classifier.T2 WI features,DWI features,and combined T2 WI and DWI(mp-MRI)features were applied to establish T2 models,DWI models,and T2-DWI models,respectively.Combined mp-MRI features and clinical parameters(age,menopausal status,abnormal vaginal bleeding and mean ADC value)were applied to establish complex multiparameter models.Predictive performance was assessed with receiver operating characteristic curve(ROC),the area under ROC(AUC),F1 score,accuracy,sensitivity and specificity.Results: In the test set,the T2,DWI and T2-DWI models based on deep learning features(AUCs range from 0.76–0.81,0.80–0.88 and0.85–0.92,respectively)outperformed the models based on radiomics features(corresponding AUCs: 0.73,0.76,0.79,respectively).Moreover,regardless of the extraction method,the complex mp model showed the best prediction performance,the AUC of the complex multiparameter model based on radiomics was 0.92,and the AUCs of the complex multiparameter models based on transfer learning ranged from 0.94 to 0.96.The Resnet50-complex multiparameter model achieved the highest AUC(0.96)and accuracy(0.87).Conclusions: Transfer learning is feasible and superior to radiomics in the differential diagnosis of uterine sarcoma and ALM in our dataset.ML models combine deep-learning features of non-enhanced mp-MRI and clinical features can achieve good diagnostic efficacy.