文献速递:肿瘤分割---- 优先注意网络,用于医学图像中多病变分割
Title
题目
Prior Attention Network for Multi-Lesion Segmentation in Medical Images
优先注意网络,用于医学图像中多病变分割
Abstract
摘要
—The accurate segmentation of multiple types of lesions from adjacent tissues in medical images is significant in clinical practice. Convolutional neural networks (CNNs) based on the coarse-to-fine strategy have been widely used in this field. However, multi-lesion segmentation remains to be challenging due to the uncertainty in size, contrast, and high interclass similarity of tissues. In addition, the commonly adopted cascaded strategy is rather demanding in terms of hardware, which limits the potential of clinical deployment. To address the problems above, we propose a novel Prior Attention Network (PANet) that follows the coarse-to-fine strategy to perform multi-lesion segmentation in medical images. The proposed network achieves the two steps of segmentation in a single network by inserting a lesion-related spatial attention mechanism in the network. Further, we also propose the intermediate supervision strategy for generating lesion-related attention to acquire the regions of interest (ROIs), which accelerates the convergence and obviously improves the segmentation performance. We have investigated the proposed segmentation framework in two applications: 2D segmentation of multiple lung infections in lung CT slices and 3D segmentation of multiple lesions in brain MRIs. Experimental results show that in both 2D and 3D segmentation tasks our proposed network achieves better performance with less computational cost compared with cascaded networks. The proposed network can be regarded as a universal solution to multi-lesion segmentation in both 2D and 3D tasks. The source code is available at https://github.com/hsiangyuzhao/PANet
在医学图像中准确分割多种类型病变与邻近组织在临床实践中非常重要。基于粗细策略的卷积神经网络(CNNs)已广泛应用于这一领域。然而,由于组织的大小、对比度的不确定性以及高度的类间相似性,多病变分割仍然具有挑战性。此外,常用的级联策略在硬件方面的要求较高,这限制了其在临床部署中的潜力。为了解决上述问题,我们提出了一种新型的先验注意力网络(PANet),该网络遵循粗细策略,用于医学图像中的多病变分割。所提出的网络通过在网络中插入与病变相关的空间注意力机制,实现了在单个网络中完成两步分割。此外,我们还提出了中间监督策略,用于生成与病变相关的注意力以获取感兴趣区域(ROIs),这加速了收敛并显著提高了分割性能。我们已在两个应用中调研了所提出的分割框架:肺部CT切片中多种肺部感染的2D分割和脑部MRI中多病变的3D分割。实验结果表明,在2D和3D分割任务中,与级联网络相比,我们提出的网络在计算成本较低的情况下实现了更好的性能。所提出的网络可以被视为2D和3D任务中多病变分割的通用解决方案。源代码可在 https://github.com/hsiangyuzhao/PANet 获取。
Methods
方法
In this section, we will go through the details of the proposed Prior Attention Network architecture. In the first part, we will offer an overview of the proposed network. We then provide details about the proposed attention guiding decoder with intermediate supervision, parameterized skip connections, and multi-class decoder with deep supervision accordingly.
在本节中,我们将详细介绍所提出的先验注意力网络架构。在第一部分,我们将提供所提出网络的概述。然后,我们相应地提供了关于带有中间监督的所提出的注意力引导解码器、参数化跳跃连接以及带有深度监督的多类解码器的详细信息。
Conclusions
结论
Multi-lesion segmentation has great significance in clinical scenarios, as multiple types of infections may occur simultaneously in a certain disease, and patients at different infection stages can have different types of lesions. For instance, ground glass opacity (GGO) and consolidation (CON.) are typical lung lesions in COVID-19 patients, where the former usually happens to early patients, and the increase of the latter may
多病变分割在临床场景中具有重要意义,因为在某种疾病中可能同时发生多种类型的感染,并且不同感染阶段的患者可能有不同类型的病变,例如磨玻璃影(GGO)和实变影(CON.)是COVID-19患者的典型肺部病变,其中前者通常发生在早期患者,而后者的增加可能
Figure
图
Fig. 1. The basic scheme of the proposed Prior Attention Network with intermediate supervision and deep supervision.
图1。所提出的先验注意力网络的基本架构,包括中间监督和深度监督。
Fig. 2. The topology of the proposed Prior Attention Network. We use 2D segmentation of COVID-19 lesions to illustrate the architecture of the proposed network.
图2.所提出的先验注意力网络的拓扑结构。我们使用COVID-19病变的2D分割来说明所提出网络的架构。
Fig. 3. The topology of the proposed attention guiding decoder
图3.所提出的注意力引导解码器的拓扑结构
Fig. 4. Visual comparison of the segmentation performance of different models on COVID-19 CT segmentation dataset. The red mask denotes the ground glass opacity and the green mask denotes the consolidation.
图4.不同模型在COVID-19 CT分割数据集上的分割性能的视觉比较。红色遮罩表示磨玻璃样不透明,绿色遮罩表示实变。
Fig. 5. Visual comparison of the segmentation performance of different models on BraTS 2020 validation dataset. The red mask denotes ET (label 4), the green mask denotes NCR/NET (label 1) and the yellow mask denotes ED (label 2).
图5.不同模型在BraTS 2020验证数据集上分割性能的视觉比较。红色遮罩表示ET(标签4),绿色遮罩表示NCR/NET(标签1),黄色遮罩表示ED(标签2)。
Fig. 6. Ball chart reporting the Dice score vs. computational complexity. The size of each ball represents the model complexity. In 2D segmentation, the balls represent the performance on COVID-19 CT Segmentation Dataset and the crosses represent the performance on CC-CCII Dataset. In 3D segmentation, the balls represent the performance on BraTS 2020 validation set.
图6.球形图表展示Dice得分与计算复杂度的对比。每个球的大小代表模型的复杂度。在2D分割中,球形代表在COVID-19 CT分割数据集上的性能,十字形代表在CC-CCII数据集上的性能。在3D分割中,球形代表在BraTS 2020验证集上的性能。
Table
表
TABLE I cross validation resul ts of 2d segmentation on covid-19 CT segmentaion dataset.the best two results are shown in red and blue fonts,respectively
表I COVID-19 CT分割数据集上2D分割的交叉验证结果。最佳的两个结果分别以红色和蓝色字体显示。
TABLE II cross validation resul ts of 2D segmentation on CC-CCII dataset. the best two results are shown in reo and blue fonts,respectively
表II CC-CCII数据集上2D分割的交叉验证结果。最佳的两个结果分别以红色和蓝色字体显示。
TABLE III ablation analysis of proposed Prior Attention Network on COVID-19 CT segmentation dataset.ds denotes the deep supervision,is denotes the intermediate supervision,AGD*denotes the attention guiding decoder without parameterized skip connections,and AGD denotes the attention guiding decoder.
表III 在COVID-19 CT分割数据集上对所提出的先验注意力网络的消融分析。DS表示深度监督,IS表示中间监督,AGD*表示没有参数化跳跃连接的注意力引导解码器,而AGD表示注意力引导解码器。
TABLE IV cross validation results of 3D segmentation on brats 2020 training dataset. the best two resul ts are shown in red and blue fonts,respectively.
表IV BRATS 2020训练数据集上3D分割的交叉验证结果。最佳的两个结果分别以红色和蓝色字体显示。
TABLE V 3D segmentation performance on brats 2020 validation dataset compared with the state-of-the-art methods methods. the best two results are shown in red and blue fonts,respectively
表V 在BRATS 2020验证数据集上与最先进方法相比的3D分割性能。最佳的两个结果分别以红色和蓝色字体显示。
TABLE VI ablation analysis of proposed prior Attention Network on brats 2020 validation dstaset.ds denotes the deep supervision,is denotes the intermediate supervision, AGD* denotes the attention guiding decoder without parameterized skip connections,and AGD denotes the attention guiding decoder
notes the intermediate supervision, AGD* denotes the attention guiding decoder without parameterized skip connections,and AGD denotes the attention guiding decoder*
表VI 在BRATS 2020验证数据集上对所提出的先验注意力网络的消融分析。DS表示深度监督,IS表示中间监督,AGD*表示没有参数化跳跃连接的注意力引导解码器,而AGD表示注意力引导解码器。