Research On Multimodal Brain Tumor Image Segmentation Method Based On Stacked Automatic Encoder

Gliomas are difficult to cure and prone to relapse.Therefore,it has always been the focus of research in the medical community.The main purpose of this subject is to achieve accurate segmentation of gliomas in different tissue areas(edema,tumor,tumor enhancement and necrosis,and normal tissue)by computer.Then,by segmenting the tumor images at different stages,the size of the tumor can be measured,the location of the lesion area can be located,and the growth of the tumor can be tracked in real time.At the same time,computer-assisted segmentation can reduce the amount of manual segmentation and make the segmentation result more objective.This thesis evaluates a fully automated method for segmenting brain tumor images from multi-modal magnetic resonance imaging volumes(MRI)based on stacked de-noising auto-encoders(SDAE.It mainly includes two innovations: First,the multimodal information is fully utilized in image classification.Based on the difference in the performance of different tissue structures in different modal images,the grayscale features of the image were extracted.Second,this paper builds a five-classification model based on the SDAE network to achieve accurate segmentation of brain glioma images.First of all,we processed the MRI images of different models by a pretreatment scheme based on top /bottom hat transformation.After pretreatment,the contrast between glioma and normal tissue areas in the image has enhanced,and the details of the image are well preserved.Then,we use the differences in information from T1,T1 c,T2,and Flair images in brain tumor images for classification.We extracted gray level patches from different models as the input of the SDAE.After trained by the SDAE,the initial network parameter will be obtained,which are adopted as an optimal initialization parameters of the feed forward neural network for classification.Then,the FFNN is fine-tuned by back propagation algorithm,and the final classification model is obtained.We use this classification model to achieve five classifications of glioma tissue regions.Finally,a post-processing scheme based on threshold segmentation to generate a mask to get the final segmentation result.Based on the automatic brain tumor segmentation framework proposed in this paper,we achieved two classifications,three classifications,and five classifications of brain tumor images on two real datasets,and achieved good segmentation results.The focus of this study is on the five classifications of glioma images.We used this five-classification model to achieve accurate division of different tissue regions of gliomas.We use the internationally accepted evaluation criteria to evaluate the segmentation results;a preliminary dice value of 86% for whole tumor segmentation has been achieved.The optimal segmentation result can reach 93%.At the same time,we compare SDAE with other classification models,which can prove that our method obtains the better performance than other state-of-the-art counterpart methods.