A Study On Automatic Segmentation Of Brain Tumor Using Deep Convolution Network In MRI Images

Glioma is a common malignant brain tumor that seriously damages the health of patients.Magnetic resonance imaging(MRI)is the main method for diagnosing and treating brain tumors.It is characterized by multi-parameter imaging,which can provide comprehensive and rich imaging information.Segmentation of brain tumor and its sub-regions from multimodal MRI images is very important.It can provide important information for the diagnosis and treatment of brain tumor,such as surgical protocol formulation,radiotherapy positioning,and chemotherapy monitoring.Manual segmentation is not only time-consuming,but also easily influenced by the physician's experience and subjectivity.Deep learning,specially convolutional neural networks,has outperformed the state of the art in many image recognition and target detection tasks in the field of computer vision.Also,CNN has excellent performance in the semantic segmentation of natural images.This provides a novel way to automatically and accurately segment the brain tumor.Unlike natural images,automatic segmentation of MRI brain tumors is a challenging task.Firstly,brain tumors can occur anywhere in the brain,have different shapes and sizes,and the boundaries are complex.Secondly,brain tumor usually presents high dispersivity and infiltration,blur cocured boundaries and poor contrast.Finally,the same brain tissue has the problem of inconsistent intensity between different patients,different modalities,and even different frames in the same modality.The purpose of the paper is using deep learning to segment both the complete tumors and its sub-regions including edema,necrosis,enhancing,and non-enhancing tumor in MRI images.In this paper,starting from the patch-based classification of MRI brain tumor automatic segmentation method,a patch-based convolutional neural network model is set up to automatic segment brain tumor.Then,for the low segmentation efficiency and poor spatial continuity of the method,a fast segmentation method of MRI brain tumor based on full convolutional network is proposed.Through experimental analysis,this method can quickly and accurately segment the complete tumors,but the segmentation of intratumoral sub-regions is not satisfactory for the imbalance between the number of pixels in tumor and norm regions.Therefore,this paper finally proposes the brain tumor segmentationbased on deep cascaded networks in MRI images.This method not only can effectively segment the complete tumors,also can significantly improve the accuracy of intratumoral sub-regions.The main work and research in the paper are as follows:(1)Starting from the idea of the patch-based classification brain tumor automatic segmentation,a patch-based convolutional neural network is constructed for the brain tumor segmentation from multimodal MRI images.Then,The model is trained by the error back propagation and the random gradient descent optimization algorithm.The experiments in Brain Tumor Segmentation Challenge 2015 database show that this model can effectively segment brain tumors,but there are large storage space,low computational efficiency and complicated process problems during prediction.(2)For the shortcomings of the automatic segmentation of brain tumor using patch-based convolutional neural network,the brain tumor segmentation in MRI images can be converted into image semantic segmentation.So,an improved method to fast and effectively segment brain tumors is proposed with introduction of the full convolutional network.This method takes into the whole image and only requires one forward propagation to obtain the segmentation result without any post-processing.Experimental analysis shows that this method can more quickly and accurately segment the complete tumors,but the segmentation results in intratumoral sub-regions are not good enough.(3)With the characteristics of full convolutional network that can quickly segment the the complete tumors,this paper proposes an improved method for segmentation of brain tumors in MRI images based on deep cascaded convolutional network.This method is divided into two-stage training.The first stage is used to locate the brain tumor area with training the full convolutional network model,and the second stage is only to segment the intratumoral sub-regions with training deep network using the small kernel.Experimental results show that this method can not only effectively segment the complete tumors,but also segment its intratumoral sub-regions.