Research On Brain Tumor Segmentation Based On Convolutional Neural Network In MRI Images

Brain tumor is a common neurological disease.The high-grade gliomas,as the most representative malignant tumor,are leading to high mortality.As a widely used imaging technology,Magnetic Resonance Imaging(MRI)has been wildly applied and has been an important part of medical diagnosis and treatment by virtue of its characteristics,such as boneless artifacts,multi-parameter imaging and radiation-free damage.Therefore,the application of MRI technology in brain tumor segmentation has become a hot topic of current research.Traditional networks mainly focus on the details of images and ignore the importance of global features in the segmentation process.In addition,the increase of the depth of network leads to a series of problems,such as over-fitting,decreasing convergence,drop-in speed and decreasing accuracy.At the same time,convolving with unified scale in the same layer fails in comprehensively representing the feature information.To address the above problems,we seek and propose an automatic segmentation method which is more effective for brain tumor structures in this study.We design a new training method——joint training of deep network and its corresponding shallow network to effectively promote the network training,which can more fully extract the image information needed for segmentation and improve convergence.In order to speed up the training process,we use the deep residual network which is widely recognized to solve some of the shortcomings of the deep network.The short connection operation is also added into the cascade layer composed of multiple convolutional layers with small convolution kernel to promote the gradient flow and speed up the training speed of the network.We add a multi-scale operation to the network to extract the both short-range and long-range contextual information with different kernels.Three variants of the proposed deep learning network are obtained by local fine-tuning in order to explore the impact of different structures on the segmentation results.Our proposal was validated in the BRATS2013 and BRATS2015 databases,obtaining the best performance and first position for the Complete and Core in Dice and Sensitivity(0.89,0.81,0.89,0.78)simultaneouslyfor the Training data set compared with five methods.Experimental results demonstrate that the ability of our proposed method in improving the training speed of network with good accuracy of segmentation.