Pushing the limits of detection with PET

正电子发射断层扫描(PET)和X射线计算机断层扫描(CT)是检测多种癌症最敏感的非侵入性医学成像手段。随着PET/CT设备和图像重建算法的不断改进，口译医生已经成为限制诊断准确性、临床吞吐量和成本的关键因素。人工智能(AI)，特别是基于机器学习(ML)的人工智能，有望帮助甚至取代医生。但人工智能在PET领域的发展一直滞后，很大程度上是由于图像数据不足，病变标记良好。我们的目标是客观地优化PET/CT用于病变检测的任务。为此，我们一直在开发将真实和准确表征的病变合成真实患者的PET/CT的方法。因此，我们可以在前所未有的规模上生成具有病变属性的准确地面真实的图像库，以推动强大的新病变检测人工智能的发展。此外，通过操纵病变属性，我们可以测试检测极限(LOD)。LOD措施可以告诉我们如何优化PET，以及如何在医生和人工智能之间进行整合，以提高对癌症的检测。这项研究计划致力于以下目标：1)开发工具，以便在当前和未来的PET系统上获取的真实患者数据中轻松生成具有良好特征的、真实的病变。2)建立参数模型，对LOD进行客观量化。3)识别和优化影响病变检测的关键因素，包括：图像重建方法、图像显示技术、人类阅读条件和图像阅读策略。4)利用超大型人工病变库开发病变检测人工智能，并将LOD作为优值因子，将人工智能与人类观察者进行比较。该研究计划利用现有的与产业界的合作，以加快研究并促进技术转移到临床实践。该计划强调与主题专家在PET模拟、PET重建、心理物理学和虚拟现实技术领域的合作。该计划将创建免费可用的合成病变图像存储库，供其他人使用，以促进创新并实现病变检测解决方案的基准，包括我们自己的。该项目的学员将被卡尔顿大学和渥太华大学的物理系和工程系录取。由于行业合作以及人工智能、ML、图像分析、模拟和虚拟现实等热门技术的使用，该项目对顶级研究生和研究生水平的研究人员具有吸引力。学员将置身于活跃的临床环境中，并将受益于与研究人员和临床医生的频繁互动。在学生和研究参与者注册期间，以及通过研究合作，强调多样性和包容性。

Hybrid positron emission tomography (PET) and x-ray computed tomography (CT) is the most sensitive non-invasive medical imaging modality for detecting many types of cancer. With ongoing improvements in PET/CT devices and image reconstruction algorithms, the interpreting physicians have become a key factor limiting diagnostic accuracy, clinical throughput and cost. Artificial intelligence (AI), especially based on machine-learning (ML), is promising to aid or even supplant physicians. But AI development in PET has been lagging, largely due to insufficient image data with well labelled lesions. Our goal is to objectively optimize PET/CT for the task of lesion detection. To that end, we have been developing methods for synthesizing realistic and accurately characterized lesions into PET/CT of real patients. We can therefore generate image libraries with accurate ground truth of lesion properties on an unprecedented scale to fuel development of powerful new lesion detection AI. Furthermore, by manipulating lesion properties, we can test the limits-of-detection (LOD). LOD measures can inform us how to optimize PET and how to integrate between physicians and AI to improve detection of cancer. This research program addresses the following objectives: 1)Develop the tools to easily generate well characterized, realistic lesions in real patient data acquired on current and future PET systems. 2)Establish parametric models for objectively quantifying the LOD. 3)Identify and optimize key factors influencing lesion detectability including: image reconstruction methods, image display technologies, conditions of the human-reader and image reading strategies. 4)Developing lesion detection AI using very large libraries of synthetic lesions and apply LOD as figure-of-merit to compare between AI and human observer. The research program leverages existing collaborations with industry, to accelerate research and to foster technology transfer to clinical practice. The program emphasizes collaboration with subject matter experts in the domains of PET simulation, PET reconstruction, psychophysics and virtual-reality technologies. The program will create freely-available repositories of synthetic lesion images for use by others to foster innovation and to enable benchmarking of lesion detection solutions, including our own. Trainees in this program will be enrolled in the departments of Physics and Engineering at Carleton University and the University of Ottawa. The program is attractive to top-tier graduate and post-graduate level researchers due to industry collaboration and use of trending technologies including AI, ML, image analysis, simulation and virtual-reality. Trainees will be embedded in an active clinical environment and will benefit from frequent interactions with researchers and clinicians. Diversity and inclusion are emphasized during student and research participant enrollment, and through research collaborations.