Detailed malaria diagnostics with intelligent microscopy

诊断疟疾的最好方法仍然是对血液涂片进行显微镜检查，以确定导致疟疾的疟原虫。这需要大约30分钟的显微镜检查，由训练有素的技术人员完成--熟练工人供不应求。该项目将创造一种智能显微镜，通过自动扫描涂片，并允许技术人员在扫描涂片后只检查平板电脑上的可疑血细胞，从而极大地提高技术人员的技能。疟疾是世界上最流行的传染病之一。它每年影响2亿人，造成大约40万人死亡--其中大多数是撒哈拉以南非洲官方发展援助国家的儿童。在减少疟疾发病率方面取得了令人印象深刻的进展，这使得对这种情况的良好诊断变得更加重要；认为每个发烧患者都患有疟疾的假设越来越不准确，这样做将浪费药物，并使潜在威胁生命的发烧得不到治疗。使用自动显微镜进行可靠、有用的诊断的关键在于计算机视觉；只需获取数字图像并将它们拼接到数字涂片中是重要的第一步，但对数字图像进行可靠的分析意味着技术人员不需要筛选许多健康细胞的图像。相反，他们可以将精力集中在算法识别出可疑特征的图像部分。一旦经过训练，我们的算法将能够识别许多寄生虫，只有在具有挑战性、模棱两可的情况下才会征求技术人员的意见，因为它无法确定地识别对象。然后，健康和感染细胞的全自动计数将允许对测试结果进行一致的量化，通知临床医生开出治疗处方，并帮助进行疾病监测。医学图像的分析提出了标准的“深度学习”方法的基本问题，即在数十万图像上训练多层神经网络。这样的算法无法准确地量化它们的不确定性(即当诊断可能不准确时标记出来)，也不能描述导致对图像进行给定分类的推理。它们需要非常大的训练数据集，这些数据集必须经常手工标记。我们将建立一个生成概率模型，虽然在大多数应用中不可行，因为可以在照片中发现的对象范围很大，但在相对受控的显微镜成像环境中是可能的。这将使我们能够给出每个细胞的概率结论，并突出显示不能可靠地归类为健康、感染或其他类型的细胞。繁殖模型还将能够识别导致分类的特征，例如在涂片的大图像中突出显示感染细胞。这两个功能都将使人们对算法有更大的信任，并允许它被用来支持而不是取代现有的临床人员，以及收集图像，从而使我们能够提高算法的性能。计算机视觉是一项强大的技术，但它需要血液涂片的高分辨率数字表示才能工作。因此，我们的项目有一个硬件组件，我们将在早期使用OpenFlexure显微镜的基础上创建一种载玻片扫描仪器，能够在现场实现血液涂片的数字化。这种仪器将使用低成本组件和台式数字制造，以便可以在当地生产-将诊所从昂贵的国际供应链中解放出来，并为当地企业家创造机会，他们建立有价值的工程和设计技能。我们已经在第一版显微镜上试验了这种方法，不久将在坦桑尼亚和肯尼亚购买，我们希望用一台全自动仪器实现更大的影响。

The best way to diagnose malaria remains microscopic examination of blood smears, to identify the plasmodium parasites that are responsible. This takes around 30 minutes of microscopy, done by a trained technician - skilled workers who are in short supply. This project will create an intelligent microscope that can greatly multiply the skills of a technician by scanning over the smears automatically, and allowing them to review only the suspicious blood cells on a tablet computer after the smear has been scanned. Malaria is one of the world's most prevalent infectious diseases. It affects 200 million per year, and causes around 400 thousand deaths - most of them children in ODA countries in sub-Saharan Africa. Impressive progress is being made in reducing the incidence of malaria, which makes good diagnosis of the condition ever more important; it is increasingly inaccurate to assume that every patient with a fever has malaria, and doing so will waste drugs and leave potentially life threatening fevers untreated. The key to reliable, useful diagnosis with an automated microscope lies in computer vision; simply acquiring digital images and tiling them together into a digital smear is an important first step, but robust analysis of the digital images means the technician need not sift through many images of healthy cells. Instead, they can concentrate their efforts on parts of the image where the algorithm identified suspicious features. Once trained, our algorithm will be able to identify many parasites, only asking for the technician's opinion in challenging, ambiguous cases when it could not identify objects with certainty. Fully automated counts of healthy and infected cells will then allow consistent quantification of test results, informing the clinician prescribing treatment and aiding in disease monitoring. Analysis of medical images raises fundamental issues with the standard "deep learning" approach of training a multi-layer neural network on hundreds of thousands of images. Such algorithms cannot accurately quantify their uncertainty (i.e. flag up when a diagnosis may be inaccurate), nor describe the reasoning that led to a given classification for an image. They require extremely large training datasets, which must often be labelled by hand. We will build a generative probabilistic model which, while not feasible in most applications due to the huge range of objects that might conceivably be found in a photograph, is possible in the relatively controlled imaging environment of a microscope. This will allow us to give a probabilistic verdict on each cell, and highlight cells that couldn't be reliably classified as healthy, infected, or something else. The generative model will also be able to identify features that led to a classification, for example highlighting infected cells in a large image of a smear. Both of these features will enable greater trust in the algorithm, and allow it to be used to support, rather than replace, existing clinical staff as well as collecting images that will allow us to improve the algorithm's performance. Computer vision is a powerful technique, but it requires high-resolution digital representations of blood smears in order to work. Our project therefore has a hardware component, where we will build on our earlier work with the OpenFlexure Microscope to create a slide-scanning instrument, capable of digitising blood smears in the field. This instrument will use low cost components and desktop digital manufacturing, so that it can be produced locally - freeing clinics from expensive international supply chains, and creating opportunities for local entrepreneurs that build valuable engineering and design skills. We have already trialled this approach with the first version of the microscope, which will shortly be available for purchase in Tanzania and Kenya, and we hope to achieve an even greater impact with a fully automated instrument.