Study On Brain Tumor Segmentation Algorithm For Magnetic Resonance Imaging

Brain tumor is a disease with high incidence rate and high mortality.Early detection of brain tumors and diagnosis of the location and shape of brain tumors are of great help to the treatment of patients.Magnetic resonance imaging(MRI)has a good imaging effect on human soft tissue,and can show different brain tumor regions.At present,the segmentation of brain tumor regions based on brain MRI is mainly manually interpreted by doctors,but it is time-consuming,labor-consuming and difficult,and different doctors determine the tumor regions according to their experience,resulting in different results.In recent years,brain tumor segmentation algorithm represented by deep learning technology has made great progress.Due to the fuzzy boundary,variable shape,small proportion in brain tissue and unstable growth position of tumor tissue,it is difficult to achieve accurate segmentation of brain tumor region.At the same time,noise,artifact and partial volume effect of MRI interfere with the automatic segmentation effect.It is a problem worthy of study to use the multi-modal features of MRI images to provide supplementary information for accurate segmentation and reduce errors.Based on this,the research contents and conclusions of this thesis are as follows:Firstly,a multi-layer cascade convolutional neural network model based on deep learning is proposed to solve the sample imbalance problem caused by the small proportion of tumor core and enhancing tumor in the image.The target of brain tumor segmentation in this study is three tumor regions with inclusion relationship in turn: whole tumor,tumor core and enchancing tumor.The cascade neural network model proposed in this thesis decomposes the complex multi classification problem into three simple two classification problems by cascading three sub modules,and completes the mapping of image spatial information of three target tasks to pixel markers in turn.The former sub module can provide multi-scale prior knowledge for the latter sub module in a multi-layer cascade way,which can soft constrain the attention scope of the latter sub module,and provide cross module auxiliary features for tumor core region segmentation task and enchancing tumor region segmentation task.Secondly,a Selective Weighted Modality(SWM)mechanism is proposed to solve the problem that the existing methods can not effectively use MRI multimodal features.Due to the different response of MRI multimodal images to different tumor sub regions,the imaging effect of multimodal images is different for different tasks.The SWM mechanism is introduced into the multi-layer cascade convolutional neural network model,so that each task can adaptively focus on learning the characteristics of high response modes according to the task characteristics,and improve the utilization of MRI multimodal information.Finally,in the experimental part,the cascaded convolutional neural network model and the weighted mode selection mechanism are verified on the Bra TS2018 dataset.The quantitative evaluation of the models and visual comparison of the results show that the proposed method can effectively improve the performance of tumor segmentation.