Research On Gastric Cancer Prognosis Prediction Method Based On Deep Learning Of Pathological Image

Among all types of cancer,gastric cancer is now one of the most well-known malignant tumors with increasing incidence and mortality rates over the years.Prognostic research of gastric cancer has always been a crucial aspect of current gastric cancer research,as accurate prognosis analysis can effectively improve the survival rate of patients.With the development of computational pathology techniques,prognosis prediction based on deep learning of pathological images has become a research hotspot.However,there is currently a lack of research on gastric cancer prognosis prediction based on pathological images,and conventional single-model,single-resolution methods have poor predictive performance.Additionally,manually labeling pathological images is difficult,making it challenging to establish highaccuracy fully supervised models.Most research only considers information under a single modality.Therefore,to address these issues and establish a highly accurate gastric cancer prognosis prediction model,this dissertation conducts research on a gastric cancer prognosis prediction method based on deep learning of pathological images.The main work is summarized as follows:(1)A gastric cancer prognosis prediction method based on the ensemble deep learning of pathological images is proposed to address challenges such as low consistency in visual review,large differences in multi-resolution images,and poor predictive performance of conventional single-model,resolution methods.Firstly,preprocessing is performed on pathological images of different resolutions for patients by segmentation and selection.Next,Res Net,Mobile Net V3,and Efficient Net V2 deep learning methods are used to extract and fuse deep features of slices under different resolutions,thus obtaining the predictive results of single-resolution sub-classifiers at the patient level.Finally,a dual integration strategy is adopted to fuse the heterogeneous sub-classifier predictive results under different resolutions to obtain the prognosis prediction results at the patient level.The method is validated using distant metastasis prediction as an example in the experiment.The experimental results show that the prediction accuracy of the proposed method on the test set is 89.10%,the sensitivity is 89.57%,the specificity is 88.61%,and the Matthews correlation coefficient is 78.19%.Compared with the single-model prediction results,the prediction performance of the proposed method has been significantly improved,which can provide an important reference for the treatment and prognosis of gastric cancer patients.(2)A semi-supervised deep learning gastric cancer prognosis prediction method based on multi-obejective optimization is proposed to address problems such as the difficulty in obtaining accurate and detailed manual pathological image labeling data in the actual clinical environment and the fact that manual selection of model hyperparameters cannot achieve optimal performance.Firstly,an unsupervised pretraining model based on the resolution sequence prediction is established using unlabelled data under two resolutions,which is then fine-tuned using supervised training.Next,the semi-supervised consistency framework is initialized using the pretraining model,and semi-supervised consistency training is performed.Finally,the multi-target Harris hawk optimization algorithm is used to optimize the hyperparameters of the semi-supervised framework to obtain the optimal predictive results.The method is validated using distant metastasis prediction as an example in the experiment.The results show that the proposed method significantly improves the predictive results compared to the fully supervised model and the prediction accuracy on the test set is 86.65%,the sensitivity is 87.03%,the specificity is 86.26%,and the Matthews correlation coefficient is 73.30%,which can provide important references for clinical auxiliary medical care.(3)A proposed multi-modal integrated deep learning method for gastric cancer prognosis prediction is presented to address issues such as the lack of clinical information guidance during image interpretation and incomplete coverage of pathological information in a single modality,which leads to poor model performance.This method is a progressive study of the ensemble deep learning-based gastric cancer prognosis prediction method proposed in Chapter 2 of this paper.The method retains the original ensemble deep learning framework and further uses LSTM to train a network for feature extraction of tabular data.Then,the deep features of image and tabular data are fused,and the prediction results are obtained through a dual ensemble strategy.The experiment validates the method by predicting distant metastasis.The results show that the prediction accuracy on the test set is 86.65%,the sensitivity is87.03%,the specificity is 86.26%,and the Matthews correlation coefficient is 73.30%and compared with the ensemble model that only uses image data,the proposed method further improves the accuracy of prognosis prediction.Finally,this article summarizes the advantages of the proposed methods in the technical field of gastric cancer prognosis prediction based on pathological images,and looks forward to future improvement plans and related research hotspots.