Abstract:

The automatic segmentation of brain tumors plays an important role in brain tumor diagnosis, surgical planning and treatment evaluation. However, it still faces huge challenges due to the high variability and fuzzy boundaries of tumors, as well as limited data and class imbalance. At present, most methods rely on manual work, which leads to time-consuming, labor-intensive and subjective influence. It is of great research significance to seek an efficient automatic segmentation method. In order to understand such methods, the research background, significance and difficulties of brain tumor segmentation are introduced, and its development process is summarized. Then the Convolutional Neural Network（CNN） based on brain tumor segmentation is described in detail from two aspects of data and structure optimization, as well as commonly used data set and indicators are introduced. Finally, the performance of the top-ranked algorithms in the BraTs challenge from 2017 to 2019 are analyzed, and the development trend of convolutional neural networks in brain tumor segmentation is discussed.

Key words: Convolutional Neural Network（CNN）, brain tumor, multi-scale, multi-task, multi-view

摘要：

脑肿瘤自动分割对脑肿瘤诊断、手术规划和治疗评估中起着重要的作用。然而，由于脑病变结构的高可变性，组织边界模糊，以及数据有限和类不平衡等问题，导致其仍面临巨大的挑战。目前，大部分分割依赖手工，耗时耗力，易受主观影响，寻求一种高效的自动分割方法非常具有研究意义。介绍了脑肿瘤分割的研究背景、意义和难点，并概述了其发展历程；从数据和结构优化两方面详细描述基于脑肿瘤分割的卷积神经网络，简介脑分割常用的数据集和性能指标；分析了2017至2019年的BraTs挑战赛中排名靠前的算法性能，并讨论分析卷积神经网络应用于脑肿瘤分割的发展趋势。

关键词: 卷积神经网络, 脑肿瘤, 多尺度, 多任务, 多视图