Prognosis Research Based On Colorectal Pathology Image Calculation

Colorectal cancer is a high risk cancer that has a strong connection with people's living habits.Colorectal cancer includes colon cancer and rectal cancer.Early detection and early treatment will greatly improve the prognosis of patients with colorectal cancer.Analysis of different types of tissue components and related histopathological features in colorectal pathology images,which play an important role in predicting postoperative survival.By predicting the prognosis of the patient after treatment,it can provide some guiding suggestions for the doctor's treatment,assist the doctor to develop a corresponding treatment plan for the patient,and make the patient's treatment more appropriate.In view of the above mentioned situation,the prognosis analysis of colorectal cancer is mainly divided into two steps.The first step is to segment the multiple tissue types in the whole slide pathological image of colorectal cancer,which can help pathologists identify different tissues more conveniently and find regions of interest more accurately.The second step analyzes the tissue components and other related features in the pathological image of colorectal cancer to establish a prognostic model,then analyzes the patient's living conditions and survival differences.In the multi tissue segmentation stage of whole slide pathology of colorectal cancer,this paper designed a deep convolutional network DeepTissue Net for segmentation of multiple tissue types.The method connects the feature maps from each convolution layer output to ensure the full utilization of feature maps.In addition,focal loss function was leaded in to alleviate sample similarity and imbalance of samples.The main purpose of this method is to segment 10 types of tissue regions in the whole slide pathological image of colorectal cancer,and multi-center data was used to verify the validity of the model.In addition,the data collected by different scanners were used to analyze different scanners impact to segmentation results.The results of comparative experiments using other different deep networks show that DeepTissue Net is more effective.Then the segmentation results are used to calculate the image features.The image features include the proportional features of tissues,deep feature of cancerous region,tissue texture features and clinical features.In addition Graph features,ClusterGraph features and Haralick features at cell level were also calcuted.All combinations of three different feature selection algorithms(mRMR,Wilkcoxon,ttest)and four classifiers(SVM,LDA,QDA,RF)were tried to predict patients' s survival status.Later,Kaplan-Meier(KM)survival curve and Log-Rank test was used for survival difference analysis.The test results on the independent test set showed that the accuracy of the model in predicting the prognosis reached 81.52% and the AUC value reached 0.77.The survival difference analysis showed that the P value was far less than 0.00001.The COX proportional hazard regression model also illustrates the significant differences in the predicted results.