Intelligent Ultrasound Imaging

超声（US）是医学成像行业发展最快的部门。最近开发的价格实惠的便携式超声扫描仪催生了许多即时超声（POCUS）应用程序，可以改变放射科以外的患者的生活。美国也在寻找新的应用，例如在肌肉骨骼方面，在扫描协议方面几乎没有智慧。其中一些涉及徒手3D US，通过该方法，US探头手动扫过感兴趣区域以获取3D图像。虽然美国技术本身已经成熟，可以解决这些新的应用，但由于缺乏熟练的美国成像操作人员，进展受到限制。由于超声检查质量在很大程度上取决于操作员的技能，因此提高可用性是向新用户和新应用部署超声成像的关键，也是本研究项目的目标。为此，我们将开发一种新的测量系统来研究探头运动和凝视模式，这些模式表征了经验丰富的超声波检查师在他们熟知或不太了解的任务中使用的视觉运动策略。我们将使用测量和机器学习方法来研究超声医师使用的与US视图质量相关的图像特征和视觉提示，以及这些视觉提示（如超声医师通常看到但未记录的中间US视图序列中所见）如何用于指导运动决策并将探头导航到给定目标。 这些新获得的知识可能会带来用于智能US成像（i-US）的新交互式导航和引导系统，从而引导经验不足的用户获取高质量的US图像。将使用受移动的机器人文献启发的新型同步定位和映射算法来研究朝向目标2D US视图的引导导航，而我们用于徒手3D US成像的自适应引导策略将基于强化学习技术的最新发展，强化学习技术是机器学习方法的一个强大子集，旨在优化顺序决策（在这种情况下，如何在检查期间移动US探头以采集期望的高质量体积数据）。我们的研究结果将在青少年特发性脊柱侧凸（AIS）的背景下得到验证，这是一种在青少年快速生长期间发生的脊柱变形，从而朝着无辐射评估和AIS随访的方向取得进展。本提案中所设想的i-US既完全新颖又非常及时，因为它可以加速新用户和新应用程序对US的采用。 事实上，i-US有可能克服成像技能的缺乏，这是广泛接受极具前景的POCUS技术和徒手3D成像协议的主要障碍，并有可能提高医疗护理标准。

Ultrasound (US) is the fastest developing sector in the medical imaging industry. The recent development of affordable and portable US scanners spawned numerous point-of-care US (POCUS) applications that can change patients' lives outside radiology departments. US is also finding new applications, e.g. in the musculoskeletal context, for which there is little wisdom available when it comes to scanning protocols. Some of these involve freehand 3D US, a method by which the US probe is manually swept over an area of interest to acquire 3D images. While US technology itself has reached the maturity to address these new applications, progress is limited by a shortage of skilled US imaging operators. Since US examination quality is highly dependent on the skill of the operator, improving usability is key to the deployment of US imaging towards new users and new applications and constitutes the objective of this research program. To this end, we will develop a novel measurement system to study the probe motion and gaze patterns that characterize the visual-motor strategies used by experienced sonographers in tasks that are well-known or poorly-known to them. We will use the measurements and machine learning methods to investigate the image features and visual cues used by sonographers that are related to US view quality, as well as how these visual cues, as seen in sequences of intermediate US views typically seen but not recorded by sonographers, are used to guide motor decisions and navigate the probe towards a given target.  This newly acquired knowledge could lead to new interactive navigation and guidance systems for intelligent US imaging (i-US) that will guide less experienced users towards high quality US image acquisitions. Guided navigation towards a target 2D US view will be investigated using novel simultaneous localization and mapping algorithms inspired from the mobile robotics literature, while our adaptive guidance strategies for freehand 3D US imaging will build on the most recent developments in reinforcement learning techniques, a powerful subset of machine learning methods that aims at optimizing sequential decision making (in this case, how to move the US probe during an examination to acquire the desired high quality volumetric data). Our findings will be validated in the context of adolescent idiopathic scoliosis (AIS), a deformation of the spine occurring during rapid adolescent growth, thereby making progress towards radiation-free assessment and follow-up of AIS. i-US as envisioned in this proposal is both entirely novel and extremely timely as it could speed up the adoption of US by new users and in new applications.  Indeed, i-US has the potential to overcome the lack of imaging skills that is the primary barrier to broad acceptance of the highly promising POCUS technology and the freehand 3D imaging protocol, with the potential to improve the standard of medical care.