[论文解读] AnatomyNet: Deep 3D Squeeze-and-excitation U-Nets for fast and fully automated whole-volume anatomical segmentation

Wentao Zhu, Yufang Huang+5 more|arXiv (Cornell University)|Aug 16, 2018

Medical Imaging and Analysis参考文献 41被引用 17

一句话总结

AnatomyNet 是一种端到端、无需图谱的3D挤压-激励U-Net，可实现对CT扫描中头颈部九个器官危及组织（OARs）的快速、全自动全体积分割。其在状态最先进方法的基础上，Dice相似系数（DSC）平均提升了3.3%，同时仅用0.12秒即可完成整个体积的分割。

ABSTRACT

Purpose: Radiation therapy (RT) is a common treatment for head and neck (HaN) cancer where therapists are often required to manually delineate boundaries of the organs-at-risks (OARs). The radiation therapy planning is time-consuming as each computed tomography (CT) volumetric data set typically consists of hundreds to thousands of slices and needs to be individually inspected. Automated head and neck anatomical segmentation provides a way to speed up and improve the reproducibility of radiation therapy planning. Previous work on anatomical segmentation is primarily based on atlas registrations, which takes up to hours for one patient and requires sophisticated atlas creation. In this work, we propose the AnatomyNet, an end-to-end and atlas-free three dimensional squeeze-and-excitation U-Net (3D SE U-Net), for fast and fully automated whole-volume HaN anatomical segmentation. Methods: There are two main challenges for fully automated HaN OARs segmentation: 1) challenge in segmenting small anatomies (i.e., optic chiasm and optic nerves) occupying only a few slices, and 2) training model with inconsistent data annotations with missing ground truth for some anatomical structures because of different RT planning. We propose the AnatomyNet that has one down-sampling layer with the trade-off between GPU memory and feature representation capacity, and 3D SE residual blocks for effective feature learning to alleviate these challenges. Moreover, we design a hybrid loss function with the Dice loss and the focal loss. The Dice loss is a class level distribution loss that depends less on the number of voxels in the anatomy, and the focal loss is designed to deal with highly unbalanced segmentation. For missing annotations, we propose masked loss and weighted loss for accurate and balanced weights updating in the learning of the AnatomyNet. Results: We collect 261 HaN CT images to train the AnatomyNet, and use MICCAI Head and Neck Auto Segmentation Challenge 2015 as the benchmark dataset to evaluate the performance of the AnatomyNet. The objective is to segment nine anatomies: brain stem, chiasm, mandible, optic nerve left, optic nerve right, parotid gland left, parotid gland right, submandibular gland left, and submandibular gland right. Compared to previous state-of-the-art methods for each anatomy from the MICCAI 2015 competition, the AnatomyNet increases Dice similarity coefficient (DSC) by 3.3% on average. The proposed AnatomyNet takes only 0.12 seconds on average to segment a whole-volume HaN CT image of an average dimension of 178 x 302 x 225. All the data and code will be available in https://github.com/wentaozhu/AnatomyNet-for-anatomical-segmentation.git. Conclusion: We propose an end-to-end, fast and fully automated deep convolutional network, AnatomyNet, for accurate and whole-volume HaN anatomical segmentation. The proposed AnatomyNet outperforms previous state-of-the-art methods on the benchmark dataset. Extensive experiments demonstrate the effectiveness and good generalization ability of the components in the AnatomyNet. Key words: Fast and fully automated anatomical segmentation, 3D squeeze-and-excitation U-Net (3D SE U-Net), radiation therapy, head and neck organ segmentation

研究动机与目标

• 为解决头颈部（HaN）癌症放射治疗计划中器官危及组织（OARs）手动勾画耗时且不一致的问题。
• 克服对仅占据几帧图像的小型解剖结构（如视交叉和视神经）进行分割的挑战。
• 处理临床数据中不一致的标注，包括由于放疗计划实践差异导致某些结构缺少真实标注的情况。
• 开发一种完全自动化、端到端的深度学习模型，消除对耗时图谱配准的依赖。

提出的方法

• 提出一种仅含单层下采样操作的3D挤压-激励U-Net（3D SE U-Net），以在GPU显存使用与特征表示能力之间取得平衡。
• 采用3D挤压-激励残差块，通过自适应重新校准通道特征响应来增强特征学习能力。
• 引入一种混合损失函数，结合Dice损失以提升类别分布的鲁棒性，以及焦点损失以缓解分割任务中的类别不平衡问题。
• 设计一种掩码损失与加权损失策略，通过仅在可用真实标注区域更新梯度的方式，处理缺失标注的问题。
• 在261个体部CT体积上端到端训练模型，所有数据均带有多种结构的标注，并以MICCAI 2015挑战赛数据集作为基准。

实验结果

研究问题

• RQ1 深度学习模型是否能在不依赖图谱配准的情况下，实现对多个头颈部OARs的快速、全自动全体积分割？
• RQ2 3D挤压-激励U-Net架构在分割视交叉等小而低可见度的解剖结构方面效果如何？
• RQ3 结合Dice损失与焦点损失的混合损失函数，在存在类别不平衡和缺失标注的情况下，能否显著提升分割精度？
• RQ4 所提出的掩码损失与加权损失策略是否能有效应对临床标注中某些结构缺少真实标注的不一致性？

主要发现

• AnatomyNet在MICCAI 2015基准测试中，相较于当前最先进方法，九个头颈部OARs的Dice相似系数（DSC）平均提升了3.3%。
• 该模型仅用0.12秒即可完成一个头颈部CT体积（平均大小为178×302×225）的分割，展现出极高的推理速度。
• 3D挤压-激励残差块显著提升了特征表示能力，尤其在视神经等小而低对比度结构上表现突出。
• 结合Dice损失与焦点损失的混合损失函数，显著提升了稀有且类别不平衡结构的分割性能，尤其在标注稀疏的结构上效果更佳。
• 掩码与加权损失策略使得模型在训练数据中某些解剖结构缺少真实标注时，仍能实现稳定且准确的训练。
• 模型展现出强大的泛化能力，经验证在具有不一致标注模式的多样化临床数据上表现良好。