关键词: 脑肿瘤, 多模态图像, 医学图像分割, 小样本分割, 深度学习

Abstract: Brain tumor Magnetic Resonance Imaging(MRI) has problems such as multi-modality, lacking of training data, class imbalance, and large differences between private databases, which lead to difficulties in segmentation. In order to solve these problems, the few-shot segmentation method was introduced, and a Prototype network based on U-net(PU-net) was proposed to segment brain tumor Magnetic Resonance(MR) images. First, the U-net structure was modified to extract the features of various tumors, which was used to calculate the prototypes. Then, on the basis of the prototype network, the prototypes were used to classify the spatial locations pixel by pixel, so as to obtain the probability maps and segmentation results of various tumor regions. Aiming at the problem of class imbalance, the adaptive weighted cross-entropy loss function was used to reduce the influence of the background class on loss calculation. Finally, the prototype verification mechanism was added, which means the probability maps obtained by segmentation were fused with the query image to verify the prototypes. The proposed method was tested on the public dataset BraTS2018, and the obtained results were as following:the average Dice coefficient of 0.654, the positive prediction rate of 0.662, the sensitivity of 0.687, the Hausdorff distance of 3.858, and the mean Intersection Over Union(mIOU) reached 61.4%. Compared with Prototype Alignment Network(PANet) and Attention-based Multi-Context Guiding Network(A-MCG), all indicators of the proposed method were improved. The results show that the introduction of the few-shot segmentation method has a good effect on brain tumor MR image segmentation, and the adaptive weighted cross-entropy loss function is also helpful, which can play an effective auxiliary role in the diagnosis and treatment of brain tumors.

Key words: brain tumor, multi-modal image, medical image segmentation, few-shot segmentation, deep learning