Research On Brain Tumor Image Segmentation Method Based On Deep Learning

Brain tumor segmentation plays important roles in radiotherapeutic planning and therapeutic effect evaluation.Due to the shape diversity,location instability,structural complexity,and the large difference in pathological symptoms for different patients,manual segmentation of brain tumors is not only time-consuming but also depending on the personal experience the physicians.Therefore,this thesis devotes itself in developing automatic methods for segmenting tumors from Brain MR images.The main work and contributions are as follows:(1)To resolve the MRI tumor segmentation problem where meaningful and important tumor feature information is less visible while noise is serious,an improved DenseNet is proposed with a novel Convolutional Attention Module(CAM)incorporated(CAM-DenseNet).Upon such an incorporation,the reusability of the DenseNet and the selectivity of the CAM are combined together.In the proposed CAM,the Mean pooling and Max pooling results from two channels are fused together for better reservation of tumor features and constructing attention coefficients,based on which effective selective weights can be obtained for low level features.The introduction of 1×1 convolution and the Leaky_ReLU activation increases the nonlinearity of the CAM and the trainability of the model.Upon defining the loss function as the sum of the multi-class Dice and the weighted cross-entropy,the imbalance between the data number for different types of tumors is alleviated.The visual analysis of the feature map and the online evaluation of the segmentation results show that the incorporation of the CAM makes the DenseNet better focuses on features containing tumor information while suppressing irrelevant ones.Furthermore,the CAM-DenseNet outperforms the U-Net and the DenseNet in brain MR tumor segmentation.(2)To consider the diversity of brain tumor shapes and sizes and also of their positions,a multi-scale DenseNet is proposed based on an improved Atrous Spatial Pyramid Pooling(IASPP)for brain MRI tumor segmentation.Here the multi-scale strategy is employed to extract more abstract tumor information from deep features.To tackle the discontinuities in the reception fields of the traditional ASPP,the sizes and distributions of the expansion rate of the dilated convolution are adjusted.Upon fusing features after Max pooling and Mean pooling,the diversity of the extracted features is improved.The visual analysis of the feature map and the online evaluation of the segmentation results show that this method can not only extract tumor features continuously but also outperform other methods in segmentation accuracy in the complete tumor area,core tumor area,and enhanced tumor area.