Deep Learning Image Enhancement for Point of Care Ultrasound

项目概要 超声因其非侵入性、非电离性和实时成像的特点而具有许多临床应用 属性。然而，超声波仍然严重依赖操作员的图像采集和解释技能。 超重和肥胖患者群体的操作技能尤其受到挑战，因为这些患者的成像伪影 例如声学杂波更加突出并降低了解剖学的显着性。为了减少 运营商面临的解释负担，我们的目标是开发一个实时的深度学习框架 声学杂波伪影抑制。 我们使用可配置的云计算工具生成初步的计算机训练数据，该工具可扩展到 8000 个 CPU 集群。该工具非常适合深度学习方法，因为它显着加快了 模拟独特的超声采集配置的周转时间可在几天内生成数据 而不是几个月。在这个项目中，我们将开源我们的云计算模拟工具，改进我们的 通过结合人体腹壁组织信息，当前的声杂波计算机数据模型 来自医学 CT 扫描的数据，并评估我们的杂波校正模型在体内数据上的性能。 为了将我们的模型结果转化为医疗提供者的解释，图像后处理是 必要的。在我们最近发表的作品 MimickNet 中，我们使用深度学习方法成功地 一些最好的临床级超声扫描仪上发现的近似后处理算法。我们 建议扩展 MimickNet 以合并后处理近似以供特定解剖用途使用 心脏和血管成像等病例。这将为研究人员提供更多现成的工具 将他们的算法研究转化为提供商熟悉的图像形式，从而简化临床翻译。 最后，便携式超声硬件的成本显着降低，使得广泛应用成为可能。 使用移动床旁超声 (POCUS)。由于许多消费设备都包含硬件 专用于深度学习应用的加速器，有机会纠正超声伪影 实时，即使受限于移动硬件。我们的初步数据表明，波束成形 操作和 MimickNet 可以在 NVIDIA P100 GPU 上以每秒 > 100 帧的速度运行。我们建议 开发一个框架，将我们的图像处理管道完全转移到移动硬件上 加速器。这项工作将使翻译新颖的图像处理算法变得像下载一样简单 软件。 我们为 POCUS 系统开发深度学习框架的工作涵盖了整个图像 从模拟数据到生成提供商熟悉的临床级图像的重建流程。这个 该框架将为以廉价且廉价的方式提高超声成像质量提供快速转化路径 广泛使用的移动硬件。

Project Summary Ultrasound has many clinical applications due to it’s non-invasive, non-ionizing, and real-time imaging properties. However, ultrasound still relies heavily on operator skills for image acquisition and interpretation. Operator skill is especially challenged in overweight and obese patient populations where imaging artifacts such as acoustic clutter are more prominent and decrease anatomical conspicuity. To decrease the interpretation burden faced by operators, we aim to develop a deep learning framework for real-time acoustic clutter artifact suppression. We generate preliminary in silico training data using a configurable cloud-compute tool that scales to an 8000 CPU cluster. This tool is ideal for deep learning methods as it significantly speeds up the turnaround time for simulating unique ultrasound acquisition configurations enabling data generation in days as opposed to months. In this project, we will open-source our cloud-compute simulations tools, improve our current in silico data model of acoustic clutter by incorporating human abdominal wall tissue information from medical CT scans, and assess our clutter correction model’s performance on in vivo data. To translate our model’s results for medical provider interpretation, image post-processing is necessary. In our recently published work, MimickNet, we use deep learning methods to successfully approximate post-processing algorithms found on some of the best clinical-grade ultrasound scanners. We propose extending MimickNet to incorporate post-processing approximations for anatomy-specific use cases such as cardiac and vascular imaging. This will provide more off-the-shelf tooling for researchers to translate their algorithmic research into image forms familiar to providers, thus easing clinical translation. Lastly, portable ultrasound hardware has significantly decreased in cost, enabling the widespread use of mobile point-of-care ultrasound (POCUS). Since many consumer devices contain hardware accelerators specific for deep learning applications, there is an opportunity to correct ultrasound artifacts in real-time, even while constrained to mobile hardware. Our preliminary data show that beamforming operations and MimickNet can run at > 100 frames-per-second on an NVIDIA P100 GPU. We propose developing a framework to transfer our image processing pipeline completely onto mobile hardware accelerators. This work will enable translating novel image processing algorithms as easy as downloading software. Our work in developing a deep learning framework for POCUS systems covers the full image reconstruction pipeline from simulated data to producing a clinical-grade image familiar to providers. This framework will provide a rapid translational path for improving ultrasound imaging quality on cheap and widely available mobile hardware.