Brain Tumor Segmentation Based On Deep Learning

Brain tumor is a disease that seriously endangers human life,health and safety.Once a patient is diagnosed with a high-grade brain tumor,he will not only suffer from pain,but also will be difficult to recover to a normal level after surgery,and the length of life will be greatly shortened.The diagnosis of brain tumors requires the use of magnetic resonance imaging to collect images of the lesion area,and manual labeling and segmentation is required,which undoubtedly takes a long time and there are certain subjective factors.The segmentation of brain tumors by computer-aided methods has been studied by a large number of researchers at home and abroad,and good segmentation results have been obtained.At present,deep learning methods based on convolutional neural networks have been widely used in brain tumor segmentation research,and have achieved certain results.However,brain tumors grow in normal brain tissue,with little difference from the borders of non-lesional areas,no fixed growth areas,and severe class imbalance.Therefore,it remains a challenging problem for network models to accurately segment different subregions of brain tumors.In the face of the above research background and segmentation problems,the work done in this paper is as follows:(1)In this paper,by improving the original U-net network model to achieve further accurate segmentation of brain tumor sub-regions,the dice coefficient has a certain increase,and the hausdorff distance is shortened.We replace some ordinary convolutions in the original U-net model with dilate convolutions,and capture more feature information through the enlarged receptive field to solve the problem of lack of computing resources.A global reasoning unit is introduced into the model,which aggregates a set of features in the MRI image globally in the coordinate space,and realizes the global reasoning of the coordinate-interaction space through the inference global pool and weighted broadcasting,which further improves the performance of the backbone architecture and realizes multiple Model brain tumor MRI images are accurately segmented.Compared with the previous segmentation results,the Dice coefficient of the enhanced tumor region increased by 1.37%,the Dice coefficient of the tumor core region increased by 3.21%,and the entire tumor region decreased by about 0.5%.Overall,good results were achieved.(2)Compared with the traditional 2D convolutional neural network,the 3D convolutional neural network has better advantages in segmenting tumor sub-regions of multimodal brain tumor MRI,and more contextual information can be extracted in the whole process.Helps segment brain tumor segments more precisely.Considering that the stacking of ordinary convolutions consumes a lot of memory,the paper uses group convolutions to alleviate this problem.Group convolution can extract more feature information under the same calculation amount and parameter amount.In order to obtain the experimental results faster,this part of the experiment uses multi-GPU technology,and uses multiple GPUs to train the model at the same time,which saves training time,increases the batch value,and increases the direction during gradient descent.Stablize.Compared with the original DMFNet network,the Dice coefficient of the tumor core area increased by about 1%,and the training time on the same device was shortened by more than 3 hours.(3)Based on React and Node,a computer-aided diagnosis and treatment system was developed,which realized the functions of computer-aided segmentation of brain tumors.It takes a long time for doctors to manually segment brain tumor subregions,but it is urgent to save the lives of patients.The computer-aided diagnosis and treatment system cannot temporarily achieve the goal of replacing the manual segmentation of brain tumors by experts.Its segmentation results can provide experts with a better reference.Combined with clinical experience,patients can be treated early,and valuable time can be saved to save patients’ lives.