基于注意力机制的脑肿瘤分割
首发时间:2018-12-14
摘要:精准的脑部肿瘤分割是癌症诊断、治疗方案和术后措施制定的重要参考因素。尽管基于全卷积网络的深度模型在脑瘤图像分割上已经取得了巨大的成功, 但是对于形态复杂、多变的肿瘤分割依然是一个巨大的挑战。本文通过观察脑肿瘤的分布,发现其具有成块出现而非弥散分布的现象。基于以上观察,本文提出一种基于注意力机制的脑肿瘤深度分割模型,模型主要包含两个网络:第一个称为注意力网络,用于检测候选肿瘤区域、同时抑制背景噪声;第二个为分割网络,用于进一步细化候选肿瘤区域的分割;这两个模块可通过监督的方式进行端到端的训练。最终,在公开的BRATS-2015脑肿瘤数据集上,本文提出算法相比传统全卷积网络有较大的提升。
关键词: 脑瘤分割 全卷积 注意力机制 膨胀卷积 BRATS挑战赛
For information in English, please click here
Attention Mechanism based Brain Tumor Segmentation
Abstract:Accurate segmentation of brain tumor is crucial for cancer diagnosis, treatment options and postoperative evaluation. Despite of the great success of previous Fully Convo- lutional Network (FCN) based methods, handling complex and varied spatial pattern of tumor tissues is still a big challenge. Furthermore, through observing the appearance of tumour, we find that kinds of tumour tissues embodyanaggregativestructure, which is a key information but ignoredby existing works. According to the regulardistribution of tumour tissues, we propose an attention-based brain tumor segmentation method, which consists of two subnets: the first subnet is used to highlight tumour areas and reduce the impact of tumour-free areas, andthe second one is designed to finely segment the tumour. Also, the whole framework can be trained end-to-end. Eventually, the extensive experiments on the public BRATS-2015 dataset demonstrate that the proposed method achieves competitive results compared with the FCN based methods.
Keywords: brain tumor segmentation FCN attention mechanism dilated convolution BRATS challenge
基金:
引用
No.****
同行评议
勘误表
基于注意力机制的脑肿瘤分割
评论
全部评论0/1000