Research And Implementation Of Brain Image Analysis Method Based On Multi-coding Structure And Cox Analysis

With the acceleration of urbanization and population aging,people are very concerned about their health.The brain is one of the core organs of the human body.Once it has problems,it will seriously endanger human’s health.Stroke and glioma are both representative brain diseases with high disability and mortality.Stroke is generally considered to be an acute cerebrovascular disease and should be diagnosed and treated as soon as the symptoms appear.CT has been the first choice for the diagnosis of stroke patients because of its speed and availability.The shape of glioma is various and its structure is complex.MRI is widely used to diagnose and treat patients with glioma,due to its advantages such as high resolution of soft tissues.At present,the assessment of stroke lesions still relies on artificial methods.Manual segmentation of stroke lesion can be very time consuming.And inconsistent evaluation standards of this method can result in different segmentation effects.Therefore,it is very meaningful to study how to apply deep learning and other artificial intelligence methods to develop automated stroke segmentation to improve the efficiency and accuracy of doctor’s diagnosis.In addition,with the increasing incidence of glioma in the population,the clinical need for predicting the survival time of patients with glioma is also becoming increasingly urgent.Survival time prediction based on MRI can not only provide necessary support for the clinical diagnosis of patients,but also help doctors formulate reasonable treatment plans.Aiming at these current clinical needs of patients with brain diseases,such as stroke and glioma,this thesis mainly studies the application of brain image analysis and processing based on multi-coding structure and Cox analysis.The main innovations and contributions of this thesis are mainly in the following:(1)A model based on multi-encoder structure for stroke segmentation is proposed.The model is elevated on the U-Net algorithm,which is a typical structure for stroke segmentation.The features extracted by U-Net is limited and its segmentation accuracy is not as expected.We add three encoders to the convolution layer of the original encoder,and fuse the feature information of multiple encoder convolutional layers during the decoding process.It can mine more features and combine more shallow features in the decoding process.In this way,improves the accuracy of segmentation.We verified this model on the ISLES018 and ATLAS datasets.The experimental results showed that our proposed model based on multi-encoder structure is superior to the traditional segmentation methods.(2)A framework based on Cox analysis for survival time of patients with glioma is proposed.We use the Pyradiomics toolkit to extract first-order statistics features;shape features,and texture features from the original MRI sequence.Then,we use Cox survival analysis to screen the extracted features.The feature subset obtained by feature selection is sent to the random forest regression algorithm to realize the survival time prediction of patients with glioma.The framework is tested on the BraTS2020 dataset.The experimental results prove that the survival time prediction framework based on Cox analysis proposed in this thesis is better than the traditional model.(3)An online diagnostic prototype system based on Flask is implemented.The system uses the Browser/Server(B/S)architecture.And the main functions realized by this system include stroke assisted segmentation,glioma assisted segmentation,glioma assisted labeling and image data management.After logging in the system,users can upload data by themselves,and the system will complete the function of segmentation or marking.Finally,the system will present the segmentation or marking results to the user in a visual form.Therefore,the system realizes the effective auxiliary diagnosis and improves the efficiency of diagnosis and treatment.