Brain Tumor Segmentation Using Multi- Cascaded Convolutional Neural Networks And Conditional Random Field

In recent years,brain tumor segmentation has become a hot issue in the field of medical image processing.Brain tumors are one of the common diseases of the nervous system and have great harm to human health.The disability and mortality are extremely high.Glioma is the most common primary tumor of the central nervous system,accounting for about 50% of all intracranial tumors.Usually,glioma is divided into high-grade glioma(HGG)and low-grade glioma(LGG),and the average life expectancy of patients who have evolved into HGG is about two years.Therefore,early detection and treatment are needed to effectively reduce the damage.Existing imaging technologies include MRI,CT,and PET,etc.Because of the unique advantages of MRI,it has become the main imaging way for the diagnosis and treatment of brain tumor.At present,many researchers have proposed efficient segmentation method for glioma MR images.But it is still challenge to segment glioma accurately due to its invasive growth and blurred boundary.In order to segment glioma effectively,convolutional neural networks(CNNs)and conditional random fields(CRFs)are used to carry out in-depth research.A method based on multi-cascaded convolutional neural networks(MCCNN)and fully connected CRFs is proposed in this paper.The main work can be summarized as follows:1)We propose a novel MCCNN architecture that allows image patches of different sizes as input to efficiently identify both the local details and contextual information.In MCCNN,three parallel sub-networks are stacked into a multi-cascaded architecture to exploit the dependencies of label and the feature of different scales.The output of the first network as additional information for the input of the subsequent network to guide the training of the subsequent network,while another concatenation is performed at the last layer of the final output to obtain smoother boundaries.2)We propose an efficient coarse-to-fine segmentation framework for glioma segmentation in multi-modal brain MRI data.Our framework mainly consists of two stages.In the first stage,the image is roughly segmented using the proposed MCCNN to obtain a probability map.In the second stage,we use CRFs to further refine the probability map.The segmentation task is transformed into an optimization problem by minimizing the energy function of CRFs.We fully consider the spatial context information to eliminate some false positives and obtain the final refined segmentation results.3)In order to verify the effectiveness of the proposed method,a series of comparative experiments are conducted on two publicly databases provided by the Brain Tumor Segmentation Challenge(BRATS).The experimental results show that the proposed coarse-to-fine segmentation method can not only achieve forefront segmentation performance,but also has low computational complexity.In addition,our method can achieve competitive segmentation results in terms of DICE coefficients and sensitivity compared to existing approaches,especially for the intractable enhanced tumor regions,which is better than most of the mainstream methods.