Brain Tumor Segmentation Methods Based On Multi-Modal MRI Images

Brain tumor segmentation is an aided diagnosis technique that separates the different tumor tissues(i.e.enhancing tumor,edema,and necrosis core etc.)from normal brain tissues: gray matter(GM),white matter(WM),and cerebrospinal fluid(CSF),for oncotherapy.Because of observing brain lesions must depend on CT or MRI technique,pathological tissues are only represented by gray value and different imaging machine may cause imaging difference,which make getting a better segmentation work to become a difficult task for traditional image segmentation methods.However,as with the development of MRI imaging technique,and multi-modal MRI imaging starts to widely spread for tumor diagnosis and treatment,tumor segmentation has rejuvenated.In such a context,based on multi-modal MRI images and machine learning methods,this paper researches following problems around brain tumor segmentation:1.Starting from the problem of tumor segmentation,against intensity inhomogeneity of MRI imaging,we proposed a difference operation by using specificity of multi-modal MRI images based on existing classical methods.And,with the help of multi-modal MRI's difference information,we developed a technique for navigating general location of tumor.So,we introduced Self-organizing Active Contour Model(SOAC),which based on Self-organizing Map(SOM)networks and Active Contour Models(ACM),to construct a Mixture Self-organizing Active Contour Model(MSOAC)based on multimodal MRI images for tumor segmentation.2.For making tumor segmentation more accurate and inspired by the performances of CNN on image classification and segmentation,we utilize convolution neural networks(CNN)and deconvolution neural networks(DeconvNet)technique to re-describe lesion structure including whole tumor region,tumor active region and tumor core region.By building specific deconvolution networks for sub-structure of tumor separately,we proposed a multi-path deep deconvolution neural networks,trained on ‘BRATS2015 training' slices by end-to-end.Different from previous approaches were used in tumor segmentation,we take slices images to train and segment named as pixel-wise,rather than patch-wise.