CAREER: Systematic Mitigation of Deep Learning Adversaries in Medical Imaging

职业：系统地缓解医学成像领域的深度学习对手

• 批准号: 2046708
• 负责人: Pingkun Yan
• 金额: $ 54.96万
• 依托单位: Rensselaer Polytechnic Institute
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-07-15 至 2026-06-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2046708&HistoricalAwards=false
• 关键词: CAREER; Systematic; Mitigation; Deep; Learning

With the enormous amounts of data being acquired by large-scale healthcare systems, computational data analysis has become an essential component in healthcare applications to process and extract information. Deep learning, a sub-category of artificial intelligence (AI), has established itself as a paradigm-shifting technology for data analytics due to its powerful ability to extract high-level data representations. However, deep learning is known to be vulnerable to adversaries, which cause algorithms to yield dramatically different results by making very small alterations to input data samples. Adversaries are particularly hazardous in medical imaging applications where an altered image may lead an AI algorithm to cause medical errors. Thus, there is an urgent need to innovate and build robust healthcare cyberinfrastructure to guard against deep learning adversaries. This project develops novel AI techniques to tackle the unprecedented challenges of adversaries in medical imaging applications from a systematic standpoint. It brings awareness to potential issues when implementing AI in healthcare and develops new tools to mitigate these issues. This research will bolster confidence in adopting AI to improve healthcare efficiency and will also attract and train the next generation of AI researchers and engineers.This project aims to develop innovative AI techniques to systematically mitigate deep learning adversaries in medical imaging applications. This project is timely as deep learning is already widely used in image reconstruction, quality enhancement, computer-aided diagnosis, and image-guided intervention and surgery. Several challenges, including detection and rectification of adversaries as well as robust algorithm training across data domains, must be resolved before achieving robust medical imaging applications. Existing methods are concerned with only the deep learning algorithms themselves and try to build universal blind robustness against arbitrary adversaries, which overlooks upstream data characteristics and downstream task specifics. This research adopts a holistic approach and is organized around a series of integrated subtopics, including detecting individual adversarial images, differentiating adversarial images from different sources, rectifying adversarial images, determining the transferability of robustness across data domains, and quantifying output uncertainties. The research will provide new insights, accurate yet robust AI techniques, and novel strategies to improve the robustness of medical imaging applications.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

随着大规模医疗保健系统获取大量数据，计算数据分析已成为医疗保健应用程序处理和提取信息的重要组成部分。深度学习是人工智能（AI）的一个子类，由于其提取高级数据表示的强大能力，已经成为数据分析的一种范式转换技术。然而，众所周知，深度学习很容易受到对手的攻击，这会导致算法通过对输入数据样本进行非常小的更改而产生截然不同的结果。在医学成像应用中，对手尤其危险，因为图像的改变可能导致人工智能算法造成医疗错误。因此，迫切需要创新和建立强大的医疗保健网络基础设施，以防范深度学习对手。该项目开发新颖的人工智能技术，从系统的角度解决医学成像应用中对手前所未有的挑战。它使人们意识到在医疗保健中实施人工智能时可能存在的问题，并开发新的工具来缓解这些问题。这项研究将增强采用人工智能来提高医疗效率的信心，也将吸引和培训下一代人工智能研究人员和工程师。该项目旨在开发创新的人工智能技术，以系统地缓解医学成像应用中的深度学习对手。深度学习已经广泛应用于图像重建、质量增强、计算机辅助诊断、图像引导干预和手术等领域，本课题的研究恰逢其时。在实现稳健的医学成像应用之前，必须解决几个挑战，包括检测和纠正对手以及跨数据域的稳健算法训练。现有的方法只关注深度学习算法本身，并试图建立针对任意对手的通用盲鲁棒性，而忽略了上游数据特征和下游任务细节。本研究采用整体方法，围绕一系列综合子主题进行组织，包括检测单个对抗图像，区分不同来源的对抗图像，校正对抗图像，确定鲁棒性跨数据域的可转移性，以及量化输出不确定性。该研究将提供新的见解，准确而稳健的人工智能技术，以及提高医学成像应用稳健性的新策略。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

Connectome transformer with anatomically inspired attention for Parkinson's diagnosis

连接组变压器从解剖学角度激发帕金森氏症诊断的关注

• DOI: 10.1145/3535508.3545544
• 发表时间: 2022
• 期刊: Computational Biology and Health Informatics
• 作者: Machado-Reyes, Diego;Kim, Mansu;Chao, Hanqing;Shen, Li;Yan, Pingkun
• 通讯作者: Yan, Pingkun

Toward Adversarial Robustness in Unlabeled Target Domains

• DOI: 10.1109/tip.2023.3242141
• 发表时间: 2023-02
• 期刊: IEEE Transactions on Image Processing
• 影响因子: 10.6
• 作者: Jiajin Zhang;Hanqing Chao;Pingkun Yan

Revisiting the Trustworthiness of Saliency Methods in Radiology AI

• DOI: 10.1148/ryai.220221
• 发表时间: 2024-01-01
• 期刊: RADIOLOGY-ARTIFICIAL INTELLIGENCE
• 作者: Zhang, Jiajin;Chao, Hanqing;Yan, Pingkun
• 通讯作者: Yan, Pingkun