基于Unet的多模态脑胶质瘤病灶自动获取模型

Automated Acquisition Model of Multimodal Glioma Lesions Based on Unet

针对脑胶质瘤分割困难,医生工作量大并且人工勾画的准确性高度依赖医生的经验,存在不能保证勾画的准确率等问题,文章提出了基于Unet的多模态脑胶质瘤病灶自动获取模型,实现对脑肿瘤的精准分割。首先根据Unet网络架构构建脑胶质瘤病灶自动获取模型,然后对不同模态的图像进行数据扩增和标准化处理后,作为网络的输入,使用dice loss来进行训练,最终得到多模态脑胶质瘤自动获取模型。实验结果显示,在Brats2017多模态脑胶质瘤数据集中,1-dice总体呈现递减,hard-dice和iou不断增加,实验效果良好。

Aiming at the problems that glioma segmentation is difficult, doctors have a large workload, and the accuracy of manual sketching highly depends on doctors’ experience, which can not guarantee the accuracy of sketching, this paper proposes a multimodal glioma lesion automatic acquisition model based on Unet to realize the accurate segmentation of brain tumors. Firstly, the automatic acquisition model of glioma lesions is constructed according to the Unet network architecture, and then the data of different modal images are amplified and standardized. As the input of the network, dice loss is used for training, and finally the multimodal automatic acquisition model of glioma is obtained. The experimental results show that in Brats2017 multimodal glioma data set, 1-dice generally decreases, hard-dice and iou continue to increase, and the experimental effect is good.