Research On Overall Survival Prediction Method Based On Digital Pathological Images

With the rapid development of imaging technology,medical imaging has become an important mode of disease management,playing an important role in disease research,screening,treatment,and prognostic evaluation.As an important basis for doctors to judge and analyze pathology,digital pathological images contain survival-related information.However,understanding pathological images through manual identification and analysis is laborious and time-consuming,and the analysis results are easily affected by the doctor's subjective factors and rely on expert experience.If ones can mine survival information from pathological images with the assistance of computers and provide prognostic information to doctors,it will greatly reduce the burden on doctors and improve the efficiency of diagnosis and treatment,which is of great significance to clinical medicine.This article studies the method of predicting the overall survival(OS)of patients based on digital pathological images.At present,most survival prediction methods use the Cox model to predict the hazard rate of patients,and then indirectly predict overall survival by estimating the baseline survival function.However,it is difficult to accurately estimate the baseline survival function,resulting in a large difference between the predicted OS and the true value.This article breaks away from the limitations of the Cox model and proposes two methods to directly predict OS.We propose an OS prediction method based on ordinal regression,which transforms the regression task into a series of binary classification subtasks.Each subtask encodes a period,learning the time series information of the data.At last,the OS is fused by the results of multiple binary classification tasks.We propose another OS prediction method based on multiple classifications,which classifies the samples according to time.The model predicts the probability that the OS of the new sample belongs to each category,and finally the predicted OS is a linear combination of the predicted probabilities.This method is easily implemented and can predict a more refined OS than classification tasks.In view of the small amount of pathological image and the censorship problem,we design loss functions that can mine the survival information of censored samples according to the proposed methods respectively,so as to make full use of the data set and improve the model's performance in predicting OS.In this paper,comprehensive experiments are carried out on the public glioma pathological image data set to verify the effectiveness of the methods,which explore the influence of censored samples,loss function and hyperparameter settings on the prediction results.The results show that both our proposed methods can predict more accurate OS than those methods based on the Cox model.