FW-HTF-R: Interpretable Machine Learning for Human-Machine Collaboration in High Stakes Decisions in Mammography

人类技术前沿项目的未来工作的具体目标是：(1)促进融合研究，采用计算机科学、工程学、学习科学、教育和劳动力培训研究以及社会、行为和经济科学的共同观点、方法和知识；(2)鼓励发展一个研究团体，致力于设计智能技术、工作组织和模式，其灵感来自于智能技术对个体劳动者、手头工作、人们学习和适应技术变革的方式、创造性和支持性工作场所（包括远程地点、家庭、教室或虚拟空间）的积极影响，以及不同规模的社会、经济和环境系统的效益；(3)促进对相互依赖的人类技术伙伴关系的更深入的基本理解，通过推进与人类工人和谐运行的智能工作技术的设计，包括考虑成年人如何学习与这些技术在工作场所相互作用所需的新技能，以及通过实现广泛的劳动力参与，包括改善身体或认知障碍人士的可及性，来推进社会需求；(4)理解、预测和探索减轻未来在人类技术前沿工作中产生的潜在风险的方法。乳腺癌是美国乃至全世界最常见的疾病和死亡原因之一。每年使用乳房x光检查的乳腺癌筛查项目在降低晚期癌症的总体负担方面非常成功。为了应对日益增加的病例量，人工智能正在广泛应用于放射学领域。到目前为止，这些人工智能系统的工作方式一直不透明，当它们出错时，放射科医生发现很难理解哪里出了问题。该项目旨在设计一种人工智能系统，该系统可以解释其推理过程，以确定女性的乳房x光检查是否包含可疑的乳房病变。该系统可以通过帮助放射科医生更好地决定是否建议该妇女进行活组织检查来改善人机交互。它还可以帮助教育医学生和其他受训者。最终，这个系统可以带来更好的病人护理，影响学术和社区临床实践。这个项目的目的不是用黑盒模型取代放射科医生：它的模型是决策辅助工具，而不是决策者，遵循放射科医生在决定是否推荐活检时必须使用的推理过程。该方法包括设计新颖的深度学习架构，该架构使用定制的可解释性定义执行基于案例的推理。与黑盒模型相比，这些模型不会失去准确性。分别的模型提出了分类的每一个乳房x线摄影任务的肿块边缘，肿块形状，和肿块密度。该项目的一个重要方面包括为放射科医生构建用户界面工具，以提供精细的注释，从而减轻混淆的有害影响。这些模型固有的可解释性将允许更好的故障排除和更容易的分析，这不仅对医学成像中的计算机辅助诊断，而且对一般的计算机视觉都将是革命性的。可解释人工智能在医疗领域的广泛应用将改变人机交互的游戏规则，并可以提高医疗保健部门的效率，不仅有助于医生管理工作量，还有助于提高患者护理质量。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。

The specific objectives of the Future of Work at the Human-Technology Frontier program are (1) to facilitate convergent research that employs the joint perspectives, methods, and knowledge of computer science, engineering, learning sciences, research on education and workforce training, and social, behavioral, and economic sciences; (2) to encourage the development of a research community dedicated to designing intelligent technologies and work organization and modes inspired by their positive impact on individual workers, the work at hand, the way people learn and adapt to technological change, creative and supportive workplaces (including remote locations, homes, classrooms, or virtual spaces), and benefits for social, economic, and environmental systems at different scales; (3) to promote deeper basic understanding of the interdependent human-technology partnership to advance societal needs by advancing design of intelligent work technologies that operate in harmony with human workers, including consideration of how adults learn the new skills needed to interact with these technologies in the workplace, and by enabling broad workforce participation, including improving accessibility for those challenged by physical or cognitive impairment; and (4) to understand, anticipate, and explore ways of mitigating potential risks arising from future work at the human-technology frontier.Breast cancer is one of the most common causes of illness and death in the US and worldwide. Breast cancer screening programs using annual mammography have been highly successful in lowering the overall burden of advanced cancers. In response to increasing caseloads, artificial intelligence is being widely adopted in the field of radiology. So far, these artificial intelligence systems have been opaque in the way they work, and when they make mistakes, radiologists find it difficult to understand what went wrong. This project seeks to design an artificial intelligence system that can explain its reasoning process for deciding whether a woman’s mammograms contain a breast lesion that is suspicious. This system can improve human-machine interactions by helping radiologists to make better decisions of whether to recommend that the woman undergo a biopsy. It can also help to educate medical students and other trainees. Ultimately, this system can lead to better patient care, impacting both academic and community-based clinical practice. This project does not aim to replace radiologists with black box models: its models are decision aids, rather than decision makers, following along the reasoning process that radiologists must use when deciding whether to recommend a biopsy. The approach includes the design of novel deep learning architectures that perform case-based reasoning with tailored definitions of interpretability. These models do not lose accuracy when compared to their black box counterparts. Separate models are proposed for each of the mammographic tasks of classifying mass margin, mass shape, and mass density. An important aspect of the project includes building user-interface tools for radiologists to provide fine annotation, which mitigates the harmful effects of confounding. The models' innate interpretability will allow for better troubleshooting and easier analysis, which will be transformative for not only computer-aided diagnosis in medical imaging but also computer vision in general. Wide implementation of interpretable artificial intelligence in the medical field will be a game changer for human-machine interaction and can improve efficiency in the healthcare sector, helping not only to manage workloads for physicians but also to improve the quality of patient care.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.