| Objective(s):According to the Global Cancer Statistics 2020 released in 2021,Breast cancer(BC)has surpassed lung cancer to become the number one cancer in women.Axillary sentinel lymph node biopsy is often performed during the frozen surgery,and clinicians choose different surgical options according to the different frozen results.Therefore,rapid pathological diagnosis of intraoperative frozen has put forward higher requirements for pathologists.HER2 is a very important indicator for breast cancer patients,not only can it predict the prognosis of breast cancer patients,but also can guide patients whether to use targeted drugs such as trastuzumab.However,in daily pathological diagnosis,the evaluation of HER2 is tedious and highly subjective.If the pathologist’s evaluation of HER2 is not accurate,it will delay the treatment of breast cancer patients or lead to excessive medical treatment.Based on the above problems,the purpose of this study is to establish a model that can accurately identify the metastases in frozen sections of axillary sentinel lymph node surgery of the breast and to build a model that can accurately predict the status of HER2 in frozen sections of axillary sentinel lymph node surgery of the breast through the combination of deep learning technology and pathomics.Methods:1.A total of 100 cases(120 intraoperative frozen sections)of axillary sentinel lymph nodes frozen in the breast cancer patients in the First Affiliated Hospital of Kunming Medical University from January 2021 to November 2022 were collected retrospectively,and grouped according to the positive and negative lymph nodes for clinical and pathological characteristics analysis.2.The Qu Path software was used to sketch the ROI area of all pathological digital slices.The outlined digital pathological sections were cut into patches with a resolution of 448×448,and all pathological patch images were normalized using Vahadane’s method.3.establish a depth learning model,and extract,regularizing and reducing that dimension of the depth learning feature.Res Net model was selected to train all the pathological patch images.4 establish model and draw model evaluation curve.The SVM,KNN,XGBoost,Light GBM and other machine learning algorithm models are used to construct the deep learning feature model.The ROC curve and DCA decision curve were drawn,and the model was evaluated.Results:1.General information: Clinical and pathological characteristics of 100 cases were analyzed.In the positive and negative groups of axillary sentinel lymph node of breast,there were significant differences in molecular typing of breast cancer,vascular cancer thrombi and nerve invasion,and HER2 status.Univariate and multivariate analyses were performed for all clinical and pathological characteristics according to the status of axillary sentinel lymph nodes.The results of the univariate analysis showed that the histological grade of the tumor and vascular tumor thrombi were significantly correlated with the metastasis of axillary sentinel lymph nodes in breast cancer patients(P < 0.05).Multivariate analysis showed a significant correlation(0.007*)between the positivity of vascular cancer thrombi and the axillary sentinel lymph node metastasis in breast cancer patients,indicating that the occurrence of vascular metastasis was prone to the axillary sentinel lymph node metastasis in breast cancer patients.2.In the research on the identification and diagnosis of metastatic lesions in frozen sections during the axillary sentinel lymph node surgery for breast cancer,we have found that the AUC under the ROC curve of Res Net50 model is the highest,which is applied to the research on the extraction of depth learning features.After the extraction,regularization and dimensionality reduction of the depth learning features,LR,SVM and KNN machine learning models are used to model the depth learning,and the results show that the AUC of the SVM machine learning model is the highest,so this model is used to draw the ROC curve and DCA decision curve of a single depth learning model.Final results show that our model AUC=0.982(95%CI0.938-1.000),sensitivity 0.889,and specificity 1.000.By comparing the single in-depth learning feature model with the in-depth learning feature+clinical feature model,the results showed that the in-depth learning feature+clinical feature model had a high AUC=0.991(95%CI 0.965-1.000),a sensitivity of 0.944,and a specificity of 1.000.The results showed that the model could accurately identify the tumor areas of frozen sections during axillary sentinel lymph node surgery of breast cancer and reduce the miss detection rate of pathologists.3.Research on predicting the state of HER2 of metastatic lesions in frozen sections during axillary sentinel lymph node surgery of breast cancer.We use Res Net50 model to train,test and verify all pathological patch images,and use the model to extract the depth learning features of all pathological patch images.After the extraction,regularization and dimensionality reduction of the depth learning features,the depth learning characteristics were modeled using KNN,random forest map,Light GBM and other machine learning algorithm models.The results showed that the AUC of the KNN machine learning algorithm model was the highest.Therefore,the ROC curve and DCA decision curve were drawn using the model.The model AUC = 0.875(95%CI 0.707–1.000),accuracy 0.833,sensitivity 0.833,specificity 0.833,true positive83%,true negative 83%,and false positive and false negative rates 17%.Conclusion(s):1.There are significant differences between the positive and negative groups of axillary sentinel lymph nodes in breast cancer in molecular typing,vascular tumor thrombus and nerve invasion,and the state of HER2.There is a significant correlation between the positive vascular tumor thrombus and the metastasis of axillary sentinel lymph nodes in breast cancer patients.2.The AUC under the ROC curve of the single deep learning feature model is 0.982,the accuracy is 0.833,the sensitivity is 0.889,and the specificity is 1.000,which has high clinical application value.3.Compared with the model constructed by single deep learning feature,the prediction model constructed by deep learning feature+clinical feature has a higher AUC=0.991,which indicates that deep learning feature+clinical feature model can accurately identify the tumor area of frozen section in axillary sentinel lymph node surgery of breast cancer and reduce the missed diagnosis rate of pathologists.4.The model for predicting the state of HER2 in frozen sections of axillary sentinel lymph node surgery was established.The AUC under the ROC curve was 0.875(95%CI 0.707 ～ 1.000),the accuracy was 0.833,the sensitivity was 0.833,and the specificity was 0.833,which indicated that the model could accurately predict the state of HER2 in breast cancer. |
PDF Full Text Request