| Breast cancer is the most common and deadliest form of cancer in women all over the world.Dynamic contrast enhanced magnetic resonance imaging(DCE-MRI)and clinical features are widely used to analyze breast cancer.Relevant studies have shown that DCE-MRI can display the physiological characteristics of breast tumors and the anatomical structure of the chest,and can provide more accurate staging information.Clinical features can reflect individual differences of patients and provide clinical manifestations related to breast cancer,while breast DCE-MRI and clinical features can be simultaneously used to study the risk of prognostic recurrence of patients,providing more accurate prediction effect for the prognosis of breast cancer.Effective prognostic recurrence risk model of breast cancer can help doctors to develop more suitable treatment for patients,and has important clinical significance.At present,most studies have evaluated the correlation between radiomics features and clinical features and the prognosis of breast cancer,but there is a lack of breast cancer recurrence risk model based on radiomics features and clinical features.In addition,the extraction and selection of imaging features based on breast DCE-MR images is a key step in the establishment of a prognostic risk model of breast cancer recurrence,and the extraction of radiomics features depends on the accurate segmentation of breast region.However,the problems of low contrast and low signal-to-noise ratio of breast DCE-MRI undoubtedly bring great challenges to the extraction of radiomics features.To solve the above problems,this paper firstly preprocessed the breast DCE-MRI,and proposed a multi-scale residual cascade U-Net segmentation model to accurately segment the breast tumor region.Then,for the tumor regions segmtioned by this segmentation model,the radiomics features were extracted.Univariate analysis,minimum redundancy and maximum correlation and random forest cross-validation were used to select the optimal subset of radiomics features.In addition,clinical features such as age,estrogen and progesterone status were obtained in this study.The radiomics features and clinical features associated with recurrence risk were ranked.Finally,a recurrence risk prediction model was established by machine learning classifier based on optimal radiomics features subset,clinical features,and combination of optimal radiomics features subset and clinical features,and the receiver operating curve(ROC)and other evaluation indicators were adopted.To evaluate the predictive value of three characteristics in identifying high,medium and low risk of relapse.Experimental data show that the segmentation index of DSC based on the multi-scale residual cascade U-Net segmentation model proposed in this paper reaches 0.7724,which can achieve accurate segmentation of breast tumors,providing an important premise for the establishment of subsequent recurrence risk model.In addition,the prediction model established by combining the combination of features and random forest classifier is the best model for predicting the risk of breast cancer recurrence.Its accuracy(ACC)is 0.884,the area under the subject operating curve AUC is 0.883,and other evaluation indicators are also at a high level.In conclusion,the prediction model proposed in this paper can effectively predict the risk level of breast cancer recurrence,and help doctors develop more personalized treatment and review programs. |
