Construction Of MSI Prediction And Prognosis Model Based On Texture Features Of Gastric Cancer Histological Images

Gastric cancer is one of the most common malignant tumors in the world.Due to the heterogeneity of cancer and the complex and diverse phenotypes of gastric cancer,the diagnosis and treatment of cancer are very difficult.Moreover,most patients with gastric cancer are already in the advanced stage of cancer when they are diagnosed,and the prognosis is poor.Therefore,the study on the heterogeneity of gastric cancer and the construction of personalized prognosis model are of great significance for the formulation of treatment plan and prolonging the prognosis of gastric cancer patients.MSI is a specific cancer phenotype caused by damaged DNA mismatch repair.Studies have shown that patients with MSI gastric cancer have unique clinicopathological characteristics and are more likely to benefit from cancer immunotherapy.MSI type usually has a good overall long-term outcomes for patients with gastric cancer,compared with non MSI patients during the same period,a higher survival rate,and with the progress of the cancer,MSI cumulative increase gradually,therefore,MSI testing in patients with gastric cancer,early diagnosis and screening of auxiliary treatment of gastric cancer clinical decision,and extend the prognosis of patients with gastric cancer has important significance.Traditional MSI tests need to be carried out by immunohistochemical analysis or polymerase chain reaction,which requires high economic and time costs and is difficult to be promoted to every patient in clinical practice.The development of histopathology and computer technology has provided a new field of vision for the diagnosis and prognosis of cancer.Therefore,this study studied the feasibility of constructing MSI state prediction and prognosis prediction model for gastric cancer patients based on histological images of gastric cancer patients.The main contents of the study include:(1)MSI prediction based on texture features of whole sections of histopathological imagesIn this study,feature extraction was carried out based on the ROI of the whole section of histomathological images of gastric cancer patients.Lasso regression was used for feature selection to select the best corresponding features and verify the correlation between features and labels.Logistic regression was used to construct a classification model,and accuracy,recall,F1 score and ROC were used to evaluate the performance of the model.(2)MSI prediction based on texture features of histopathological imagesIn order to realize MSI detection of gastric cancer patients at a lower cost,quantitative image features were extracted based on easily obtained histopathological images of gastric cancer patients.The features most related to MSI status of gastric cancer patients were selected by feature selection technology.Based on the features obtained,five machine learning classification models were trained at the slice level,The hard voting method is used to predict the MSI state at the individual level based on the prediction results at the slice level.Aiming at the limitations of the hard voting method,a weighted voting algorithm is proposed to predict the MSI state at the individual level.The performance of the classification model is compared and analyzed at the slice level and the individual level through the accuracy,recall,F1 score,ROC and other indicators.(3)Establishment of prognosis prediction model for gastric cancer based on MSI predictionIn order to build the prognosis of gastric cancer patients with personalized predictive model,improve the performance of the conventional prognostic model,research based on the prediction of gastric cancer patients with MSI,build the MSI risk score for each gastric cancer patients,patients with gastric cancer based on MSI risk score divided into high-risk and lowrisk groups,using Kaplan Meier-survival in patients with gastric cancer curve drawing analysis,combined with the Log-Rank test compares the differences between high and low risk group of patients with gastric cancer survival probability,using multi-factor Cox regression to explore the MSI risk score,TNM staging and clinical information’s influence on the prognosis of patients with gastric cancer,Five prognostic prediction models based on different prognostic factors were constructed,and their prognostic performance was analyzed and compared by using C-index,calibration curve and decision curve.In this study,the quantitative characteristics of Histopathologic image of gastric cancer patients were extracted,and the MSI status prediction was realized by machine learning algorithm.Furthermore,the MSI risk score was constructed for each patient based on the MSI status prediction,the relationship between the MSI risk score and the prognosis of gastric cancer patients was analyzed,and a prognostic prediction model was built based on the MSI risk score.The results show that the proposed method can achieve the prediction of MSI in gastric cancer patients at a lower cost,and the prediction model combined with MSI risk score has better prognostic performance compared with the traditional prognostic model,which can provide a better reference for prognosis judgment and clinical decision making of gastric cancer patients.