Research On Survival Prediction Of Glioma Patients Based On Multimodal MRI Image Analysis

With the continuous development of medical imaging technology and the extensive combination of artificial intelligence and medical diagnosis,it is of great research value to construct machine learning models for prognostic survival prediction of glioma patients based on MR image data.However,related studies in this field generally suffer from some problems:the constructed high-dimensional image feature space has large redundancy and the designed prediction model has poor generalization ability on unknown data,so it is difficult to achieve the desired prediction effect.To alleviate the above problems,this study designs a method to optimize the multi-sequence MR image data from two perspectives of reducing the multisequence image space and compressing the high-dimensional radiomic features.Furthermore,a hybrid stacking ensemble model based on the two-stage calculation mechanism is constructed in this study for prognostic survival prediction of glioma patients.The main works are as follows:(1)For multi-sequence MR images,this study extracts the image features of multiple tumor sub-regions under different MR sequences as our initial research data.According to the characteristics of MR image data and glioma tissue structure,this study designs a multisequence MR image data optimization method based on the association rule of image sequence and lesion subregion,which evaluated and optimized the potential multiple MR sequence combinations to reduce the scale of the MR image feature space.The image data space determined by the proposed optimization algorithm has a smaller size and less redundant information,and provides a better data base for building prediction models.(2)A clustering-based image feature compression algorithm is proposed in this study to optimize multi-sequence MR image data and classify patients into two types of survival stages.The obtained compressed features are the highly semantic and condensed representation of the original image feature set,which can perform the sequence optimization task more efficiently and accurately,and achieve high precision survival stage classification for all patients.(3)This study constructs a targeted two-stage prediction calculation method by first classifying patients into two different survival stages and then predicting prognostic survival days separately.In this mechanism,two differentiated sub-prediction models based on glioma patients with different survival stages are constructed to learn the data characteristics better.A dedicated hybrid ensemble prediction model based on stacking architecture is constructed for long-survival stage patients and has better performance than the single basic prediction model in this study.(4)To verify the effectiveness of the proposed method,this study conducts experiments on the Bra TS Challenge 2020 public dataset.In the MR sequence optimization experiments,the effects of feature sets constructed from different sequences are compared.In the clusteringbased feature compression experiments,a comparative analysis is performed with several representative feature reduction algorithms.In the experiments of constructing the prognostic survival prediction model,the present method is compared with several mainstream prediction methods,and designs several sets of comparison experiments to verify the superiority of the proposed hybrid stacking ensemble model based on the two-stage calculation mechanism in this study.In the study of prognostic survival prediction of glioma,the proposed feature set optimization method in this thesis has effectively utilized multiple sequences of MR imaging data for prediction calculation,while a targeted ensemble model is constructed for survival prediction.Experimental results show that the method proposed in this study significantly reduces the prognostic survival prediction error.