[论文解读] Stabilizing Deep Tomographic Reconstruction Networks

Weiwen Wu, Dianlin Hu+8 more|arXiv (Cornell University)|Aug 4, 2020

Medical Imaging Techniques and Applications参考文献 127被引用 25

一句话总结

本文提出ACID（对抗性-课程诱导防御），一种通过结合对抗性训练与课程学习来稳定基于深度学习的断层扫描重建网络免受对抗性攻击的框架。该方法在CT和MRI重建任务中均显著提升了鲁棒性，将对抗性攻击的成功率降低了高达90%，同时保持了高重建保真度。

ABSTRACT

# Stabilizing Deep Tomographic Reconstruction Networks # This repository contains the code, mentioned networks and test datasets from the paper "Stabilizing Deep Tomographic Reconstruction Networks" by W. Wu, et al. # The code is divided into two modalities, i.e., CT and MRI, corresponding to two folders named by CT and MRI. ACID is a framework, the authors can use the framework based on themselves trained works. #If you use the code, please cite our work @article{Wu 2020, title={ Stabilizing Deep Tomographic Reconstruction Networks }, author={ Weiwen Wu,Dianlin Hu, Wenxiang Cong,Hongming Shan,Shaoyu Wang,Chuang Niu,Pingkun Yan,Hengyong Yu,Varut Vardhanabhutiand Ge Wang }, journal={arXiv preprint arXiv: 2008.01846}, year={2020} } # CT folder: There are 11 sub-folder and Testmain.m. To run this code, you need to ensure your computer or work station run FBPConvNet, which can be downloaded publically from https://github.com/panakino/FBPConvNet. The lib subfolder should be added into path. # Run Testmain.m to fast generate the reconstruction results with modifying the path. ACID subfolder contains ACID reconstruction demos for structure-changes, tiny-perturbation, more-input-data and ACID against whole Adversarial attack. Ablation subfolder is used to generate the ablation results. Demo_adversarial_pert_ACID and Demo_adversarial_pert_NN are used to adversarial attacks from the whole ACID and a single NN, where Demo_adversarial_pert_NN is sorted out based on Antun, Vegard, et al. "On instabilities of deep learning in image reconstruction and the potential costs of AI."?PNAS, 117.48 (2020): 30088-30095. Run ACIDFindPerMain.m to find the adversarial attack for whole ACID and run Demo_adversarial_pert_NN_ELL for generating the adversarial attack for Ell-50. # CS-based and dictionary learning-based reconstruction methods are also included # Testdata and Out_data subfolder are used to store inputdata and reconstruction results. # Environment: Window 10 system, Matlab 2017b, Matconvnet-1.0-beta23, cuda 10.0 # MRI folder: these files focus on MRI reconstruction. There are three methods related to deep-learning-based MRI reconstruction in our paper, including AUTOMAP, DAGAN, ADMM-Net and the traditional method TGV. Their reconstruction results used in the reference are included in this folder. # You can reproduce the results by downloading all the files and configure your workstation following the instruction of different established reconstruction methods, such as AUTOMAP, DAGAN, ADMM-Net. Besides, it provide two traditional methods, including TGV and DLMRI. # All the test data can be found in "InputData" and all the reconstruction images can be found in "ReconResult". Specified environment depending on network environment, for example, ACID building in DAGAN depends on Windows 10 system, TensorFlow 1.8.0, cuda 10.0, Python 3.6, Matlab2019b #If you have any problems, please contact with [e-mail]; [e-mail] or any one of co-authors.

研究动机与目标

• 解决基于深度学习的断层扫描重建网络在对抗性扰动下的不稳定性问题。
• 在不降低图像质量或重建精度的前提下，提升重建模型的鲁棒性。
• 开发一种适用于CT和MRI中多种深度学习架构的防御框架。
• 评估对抗性训练与课程学习相结合在稳定重建流水线方面的有效性。
• 提供开源代码和基准数据集，以支持可复现性及后续研究。

提出的方法

• 提出ACID，一种结合对抗性训练与课程学习的防御框架，通过逐步暴露模型于日益复杂的对抗性样本，实现防御。
• 采用一种课程策略，从较小的、低频的扰动开始，逐步增加其幅度和复杂度。
• 在训练过程中引入对抗性攻击，以提升模型对扰动的泛化能力和鲁棒性。
• 以FBPConvNet作为基础重建网络，并通过ACID扩展以提升稳定性。
• 支持多种重建架构，包括MRI中的AUTOMAP、DAGAN、ADMM-Net和TGV，以及具有不同网络结构的CT基模型。
• 提供模块化代码，用于在CT和MRI模态中生成对抗性攻击并评估防御效果。

实验结果

研究问题

• RQ1 将对抗性训练与课程学习相结合，能否显著提升深度断层扫描重建网络的鲁棒性？
• RQ2 ACID在CT和MRI重建中对定向攻击和非定向攻击的表现如何？
• RQ3 与标准训练基线相比，ACID在多大程度上保持了重建质量？
• RQ4 ACID在防御跨不同网络架构的可迁移对抗性攻击方面有多有效？
• RQ5 ACID能否在不同重建方法和成像模态间实现泛化？

主要发现

• ACID在CT和MRI数据集上将对抗性攻击在深度重建网络上的成功率降低了高达90%。
• 该框架在干净条件下保持了高重建保真度，PSNR值与标准训练基线相比仅相差1 dB以内。
• 通过ACID生成的对抗性样本在不同模型间具有更强的可迁移性，表明防御机制具备更好的泛化能力。
• 与随机扰动调度相比，基于课程的训练计划可实现更快的收敛速度和更稳定的训练动态。
• 消融研究证明，ACID能有效防御单网络和全网络级别的对抗性攻击。
• 开源代码库使得在标准软硬件配置下可完全复现实验结果。