Tumors have always been a feared disease, and brain tumors have an incidence rate of 1.5% and an alarming 3% mortality rate in the population, and have been feared because of their extremely high incidence and mortality rate[1]. Brain tumor is a cancer of the brain tissue, which is formed when the brain tissues become cancerous or metastasize to other tissues in the skull. With the development of medical imaging technology, imaging technology has gradually been applied to tumor detection. At the beginning, Computed Tomography (CT) technology was used for detection, but with the development of magnetic resonance physics, and then combined with the theory and technology of digital image reconstruction, Magnetic Resonance Imaging (MRI), is slowly taking shape, because it has no ionizing radiation damage to the body and many imaging parameters have gradually become the mainstream of medical detection of brain tumors[2]. However, most of the current clinical diagnosis of brain tumors is based on the clinician's experience. The manual method to segment, diagnose and annotate the tumor images is inefficient and demanding for image analysts, and it is easy to miss the best treatment time of patients [3]. Therefore, how to efficiently diagnose brain tumor images and reduce the diagnostic error of images has become a research direction for many researchers. Currently, deep learning-based intelligent algorithms are widely used in brain tumor analysis tasks, and CNNs are adopted by researchers for their good segmentation performance and the convenience of feature extraction[4]. However, CNN is prone to computational redundancy when processing a large number of dense images [5]. Therefore, FCN[6], U-Net [7] and other derived algorithms based on CNN have been proposed. However, many brain tumor segmentation algorithms still have many problems, such as the segmentation accuracy and recognition accuracy of the algorithms are not high enough, and the attention to details is not enough. In this paper, we propose an improved segmentation network based on U-Net to solve these problems, using a tandem encoding-decoding model and proposing a new loss function to increase the weight of samples that are difficult to segment classify. The experimental results show that it outperforms several other commonly used derived models based on CNN in terms of segmentation performance and tumor recognition accuracy.
How to perform image segmentation is a key problem in the field of Computer Vision(CV), and image segmentation generally includes semantic segmentation and instance segmentation [8]. The brain tumor segmentation in this paper is using semantic segmentation. The evaluation of the segmentation ability of the semantic segmentation model needs to focus not only on the overall segmentation of the segmented image, but also on the edge segmentation of the segmented image. So, how to design the segmentation algorithm becomes the most important, and different researchers have proposed different methods for the research of segmentation algorithm. With the rise of neural network models and the development of deep learning, segmentation networks based on deep learning have been rapidly developed and applied. Starting from the concept of neural network proposed by Le Cun, neural networks have been developed rapidly, and various neural network structures began to emerge slowly, such as AlexNet [9], VGG[10], ResNet [11], etc. Although these networks have advantages in the field of image recognition and prediction, the advantages in accurate semantic segmentation of images are not so obvious. In order to change this situation, Shelhamer et al. proposed FCN, applying FCN to semantic segmentation of images[6]. They achieved the segmentation mainly by replacing the fully connected layers of net with convolutional layers, and the results showed that the image semantic segmentation did outperform the other Convolutional Neural Networks (CNN). The reason is that full convolutional networks (FCNs) require high data volume, and such brain tumor images are relatively few and precious in medicine. To solve this problem, Ronnerberger et al. modified the full convolutional network by adopting transposed convolution, up-sampling, and fusion of context features and detail features to form U-Net, which can obtain enough data features with few brain images, and the segmentation effect is significantly better than that of full convolutional network (FCN). But there are still problems of incomplete information and low segmentation accuracy when performing brain tumor segmentation. In order to solve the remaining problems of U-Net network, Alom proposed a recursive neural network based on U-Net and a recursive residual convolution neural network based on U-Net model.[12]. Zhang et al. Used U-Net extended path to design residual connection, and proposed a depth residual U-Net for image segmentation[13]. Milletari proposed a 3D U-Net model, which uses 3D convolution kernel to expand the original U-Net structure, and then adds residual units to further modify the original U-Net structure [14]. Salehi uses auto context algorithm to enhance U-Net to improve segmentation effect[15]. Zhou et al. Used the nesting method to replace the original connection method [16]. Wanli Chen, Yue Zhang et al. proposed a stacked U-net with a bridging structure to address the problem of increasing training difficulty as the number of layers of the network increases [17]. The above segmentations model can only segment images, but cannot grade segmented tumor. To achieve this clinical need, Mohamed A. Naser and M. Jamal Deen first used the trained segmentation model and MRI images for mask generation, and then used a densely connected neural network classifier to classify the tumor [18].