Afshar Parnian, Plataniotis Konstantinos N, Mohammadi Arash
Annu Int Conf IEEE Eng Med Biol Soc. 2020 Jul;2020:1075-1079. doi: 10.1109/EMBC44109.2020.9175922.
Brain tumor is among the deadliest cancers, whose effective treatment is partially dependent on the accurate diagnosis of the tumor type. Convolutional neural networks (CNNs), which have been the state-of-the-art in brain tumor classification, fail to identify the spatial relations in the image. Capsule networks, proposed to overcome this drawback, are sensitive to miscellaneous backgrounds and cannot manage to focus on the main target. To address this shortcoming, we have recently proposed a capsule network-based architecture capable of taking both brain images and tumor rough boundary boxes as inputs, to have access to the surrounding tissue as well as the main target. Similar to other architectures, however, this network requires extensive search within the space of all possible configurations, to find the optimal architecture. To eliminate this need, in this study, we propose a boosted capsule network, referred to as BoostCaps, which takes advantage of the ability of boosting methods to handle weak learners, by gradually boosting the models. BoosCaps, to the best of our knowledge, is the first capsule network model that incorporates an internal boosting mechanism. Our results show that the proposed BoostCaps framework outperforms its single capsule network counterpart.
脑肿瘤是最致命的癌症之一，其有效治疗部分依赖于肿瘤类型的准确诊断。卷积神经网络（CNN）在脑肿瘤分类方面一直处于领先地位，但它无法识别图像中的空间关系。为克服这一缺点而提出的胶囊网络对复杂背景敏感，无法专注于主要目标。为解决这一不足，我们最近提出了一种基于胶囊网络的架构，它能够将脑部图像和肿瘤粗略边界框作为输入，从而获取周围组织以及主要目标。然而，与其他架构类似，该网络需要在所有可能配置的空间内进行广泛搜索，以找到最优架构。为消除这一需求，在本研究中，我们提出了一种增强胶囊网络，称为BoostCaps，它利用增强方法处理弱学习器的能力，通过逐步增强模型来实现。据我们所知，BoosCaps是第一个纳入内部增强机制的胶囊网络模型。我们的结果表明，所提出的BoostCaps框架优于其单胶囊网络对应物。
