Development of machine learning techniques for accessible and inexpensive imaging of COVID-19 with ultrasound

意大利急诊室(ER)研究人员和中国最近发表的论文显示，与目前使用计算机断层扫描(CT)或胸部X光检查的做法相比，使用超声波对肺部进行成像具有几个优点。这些论文表明，可以使用B线检测新冠肺炎感染，B线是由于新冠肺炎感染而在超声图像中产生的水平线。 Clarius超声仪可作为廉价的筛查工具用于新冠肺炎试验舱、分诊前和床边监测患者。它们还可以用于对因辐射暴露而无法接受CT或X光成像的孕妇进行成像。最后，Clarius超声设备比CT扫描仪便宜200多倍，这可以大大节省医疗成本。 尽管有这些吸引人的功能，但超声波仍然依赖于用户，需要熟练的超声波技师来收集高质量的图像。此外，检测超声图像中的B线需要由专业超声医师仔细检查图像，这进一步限制了该筛查工具的可及性。鉴于急诊室目前缺乏医疗专业人员，这一伙伴关系有两个目标来解决这些问题： 1.当新手用户移动和旋转胸壁上的探头时，通过自动选择最佳超声图像来简化图像收集。 2.通过自动检测和定位最佳帧中的B线，简化了图像的解释。 在与Clarius的密切合作下，我们将开发解决这些问题的新型神经网络(NN)。我们将制造模拟健康和感染肺部的模型来训练和测试我们的神经网络。我们将利用机器学习的最新进展来开发神经网络，用于处理具有不同图像设置的不同患者的未见数据。

Recent papers from Emergency Room (ER) researchers in Italy and China has shown several advantages of imaging the lungs with ultrasound instead of the current practice of using Computed Tomography (CT) or chest X-ray. These papers showed that COVID-19 infection can be detected using B-lines, horizontal lines in the ultrasound image that are created because of the COVID-19 infection. Clarius ultrasound devices can be used as inexpensive screening tools in COVID-19 test-pods, for pre-triage, and in bedside for monitoring patients. They can also be used for imaging pregnant women who cannot undergo CT or X-ray imaging due to radiation exposure. Finally, Clarius ultrasound devices are more than 200 times less expensive than CT scanners, which can substantially save healthcare costs. Despite these attractive features, ultrasound remains user-dependent, and requires a skilled sonographer to collect high-quality images. In addition, detecting B-lines in ultrasound images entails careful inspection of images by an expert sonographer, which further limits accessibility of this screening tool. Given the current shortage of medical professionals at ERs, this partnership has two goals to address these issues: 1. Simplify image collection by automatically selecting the best ultrasound image as a novice user moves and rotates the probe on the chest wall. 2. Simplify interpretation of images by automatically detecting and localize B-lines in the best frame. In close collaboration with Clarius, we will develop novel Neural Networks (NN) that address these problems. We will fabricate phantoms that mimic healthy and infected lungs to train and test our NN. We will exploit latest advances in machine learning to develop NNs that work for unseen data from different patients with different image settings.