Description: On the road of life, have you ever felt thrilling race against time?
On the road of life, have you ever felt thrilling race against time?
In fact, many diseases if "early check early detection", it can be nipped in the bud period.
Alibaba AI learned to cardiovascular medical identification technology . It accurately extracted from coronary artery CTA image, the efficiency of nearly a hundred times higher than the conventional method, only 0.5 seconds to complete testing.
Medical diagnostic imaging direction is a popular application of AI medical, diagnostic imaging abnormalities broad spectrum of diseases throughout the liver, lung, bone, breast, thyroid, heart and other organs, which is a recognized cardiovascular difficult areas, few players can reach.
Before arriving in the heart, Ali has hit a lung nodule detection, world-class breakthrough liver nodules diagnostic techniques. Two years from the lungs, liver to cardiovascular "triple jump" so Ali AI AI advanced medical imaging in the field of "home run player."
Ali AI growing at the rate of visible towards the "whole organ" identified the finish line of evolution, Ali lung nodule detection, diagnosis and treatment of orthopedic aid and other new technologies have been landed commercial.
Once the disease spectrum lateral and longitudinal ground is formed breakthrough technology, drive technology will change AI medical imaging department.
No radiological no modern medicine
No radiological no modern medicine, before Roentgen discovered X-rays, we can not see their bones through flesh and organs - if you do not consider the words of human anatomy.
Now, the typical picture of the work appears in the Department of Radiology hospitals around the world, a group of doctors looking at a bunch of computers, thousands of doctors from across the fundus image every day. In China's major hospitals, radiologists and computer on each working day, as the time is often more than 10 hours. The group is the hospital's vision depression.
People are made of meat, a large number of mechanical repetition action consumes energy and intelligence, has become one of the causes of misdiagnosis and missed diagnosis.
New technological changes are occurring, artificial intelligence has sounded the door of modern medicine. Global generates annual trillion GB magnitude of medical imaging, machine-assisted medical treatment if we make the film, how many intellectual resources will be liberated, doctors can do more important things, such as pay would give the patient a little time.
This is a "Mount Hope ran a dead horse," the vision, looked far off to know how hard it is.
For individual organs, individual diseases, AI can be used for single-point breakthrough yet, but if you want to win across the board for all types of organs, data, algorithms, calculate power requirements to be promoted several orders of magnitude.
Meanwhile, medical diagnostics is a system of self-consistent, procedures are complicated and not without pride links, new technology to be embedded in the meantime, is facing more difficult than other scenes floor level.
Why get together medical AI lung nodule areas?
Lung nodule detection is the most familiar medical imaging AI territory.
When the tumor may be just the beginning of a nodule. However, many patients with lung cancer at the time of initial medical treatment, the resulting sentence is late.
The number one fighting cancer, the best early detection, early diagnosis, early treatment, can not be easily detected pulmonary nodules, early nodules mostly less than 10mm, generally do not lead to significant discomfort, a lot of people missed the best treatment period .
Compared to the number of potential patients, imaging department can digest of cases far less than social needs. Shooting chest CT screening for lung nodules, the number of CT images of each case more than 200, a doctor image processing up to dozens of cases every day. Fatigue strength of war, manual errors inevitable. This is the ideal value of artificial intelligence to play a scene.
July 2017, Ali AI break the world record in the international authority of the lung nodule detection contest LUNA16, with an average of 89.7% recall rate (the proportion of successes in the sample data in the proportion of nodules found) win.
LUNA16 official website ranking
Competition requires teams to find lung nodules in the lung CT 888 parts of the sample, the sample included 1186 lung nodules, more than 75% are less than 10mm. Ali AI whole without human intervention, automatically read the patient's CT sequence, to direct the output of the detection of pulmonary nodules.
2017 the "Year of pulmonary nodules," said the AI industry - currently known most AI lung nodule detection technology breakthroughs have occurred in this year. Today, the domestic companies alone have at least dozens claimed to have achieved a lung nodule detection algorithm.
Detection of pulmonary nodules in medical imaging field AI entry level, there are two reasons threshold algorithm, first image relative pulmonary nodules "readable", intuitive image screen, less disturbances, wherein a pattern; and secondly pulmonary nodules related to public data and more, get convenient, low cost training machine.
Unfortunately, For many medical AI, the lung nodule is both the starting point is the end.
From the liver, the lungs to the cardiovascular, Ali AI "triple jump"
Want to get through medical imaging disease spectrum, we must reposition itself in hard currency - algorithm.
2017 年之后，阿里 AI 继续高速奔袭，连续拿下肝结节诊断和心脏冠脉提取的两项世界顶级赛事冠军，宣示了在算法领域无可匹敌的优势。
2018 年 12 月，阿里AI从近百支队伍中脱颖而出，在全球LiTS(Liver Tumor Segmentation Challenge，肝脏肿瘤病灶区CT图像分割挑战)获得两项第一。
肝脏是人体管状物分布最密集的器官，内含门静脉、肝静脉、肝动脉、胆管系统等多套管状系统。肝结节形态多样，结节间灰度分布差异大，与周围组织灰度相似甚至没有清晰的边界，对AI的“视力”挑战大于肺结节。
阿里 AI 通过对 CT 图像层间信息和层内信息融合的网络结构分析解决了肝结节类别多样性的问题，用到了基于原子卷积的空间金字塔池化（Atrous Spatial Pyramid Pooling）、亚像素卷积（Sub Pixel Convolution）等技术。目前，阿里团队正进一步研究如何判断肝结节的良恶性。
半年后，在 2019 年的心脏冠脉中心线提取鹿特丹比赛（Rotterdam）上，阿里 AI 获得全自动提取赛事第一名，相关论文被国际顶级医学影像会议MICCAI 2019 提前接收。
从CTA影像中准确提取心脏冠脉中心线是冠心病影像诊断的必备条件。通常的流程是，医生根据二维图像对血管进行三维重建，形成曲面重建视图，手动提取冠脉血管，寻找并标注血管斑块，判断血管病变性质，从而确定治疗手段。
心脏冠脉平均长度164mm，几何特性复杂、血管特别细小。比起肺结节的静态扫描图像，为不断跳动的心脏作三维图像重建棘手得多。传统的心脏冠脉中心线提取方法大多存在人工交互多、耗时长等缺点。
阿里AI提出了判别式冠脉追踪模型，三维卷积神经网络构成的模型，充分利用三维空间特征，从影像中迭代搜索完整血管，无需人工交互，提取单根冠脉血管平均耗时 0.5 秒，提取完整冠脉树用时不超过 20 秒，相比传统方法效率提升近百倍。
心血管疾病诊断的复杂性，导致AI医学影像识别在该领域应用极少。阿里的技术突破，让AI辅助医生进行心血管疾病诊断的未来变得近在咫尺。
从肺、肝到心血管的“三级跳”，姿态有多轻盈，算法就有多厚重。算法突破没有捷径，拼的就是人才浓度。事实上，阿里达摩院已经悄然聚齐国内最强 AI 研发者阵营——达摩院现有 10位IEEE Fellow、20 多位知名大学教授，一半以上科学家拥有名校博士学历。
夺冠以后，阿里 AI 还做了这两件事
光有算法远不足以推动技术落地。算法模型与现实场景的结合才是难点，具体到医疗 AI ，需要解决的问题很多，比如，如何在真实医疗环境中证明模型的准确率，如何解决真正的临床关切问题，如何确立服务模式和商业模式。
比起 LUNA16 夺冠，阿里在 2017 年做的另外两件事，更能显示其打法思路。
当年 3 月底，阿里云联合英特尔、零氪科技发起第一季“天池医疗 AI 大赛”， 以肺结节智能识别和诊断为课题，开展肺癌早期影像诊断的应用探索。16 家三甲医院的医师组成专家指导团，来自 20 个国家的2887 队伍报名竞技，整个赛程长达半年，部分优秀算法最终转化为了实际解决方案。
三个主办方各司其职，阿里云提供机器学习训练平台，单点支持数百 GB 内存，每次迭代可高速处理 32 张以上 128x128x128 甚至更大规模的 3D 图片；英特尔提供由第二代强融核处理器打造的高性能计算集群，保障高强度算力供应；零氪联合 16 家医院提供全球最大规模的 2000 份“科研级胸部CT数据集”。
这场非商业赛事动员了整个行业的智慧，推动肺结节检测的算法优化和技术沉淀。但它更重要的价值在赛事榜单之外——让专业医师与算法工程师走到一起，探讨影像学与 AI 的协作方式，从算法层面就关注 AI 嵌入医学流程的可行性。
“开放”一直是阿里追求的 AI 产业生态。阿里的 AI 平台能为中小企业提供人工智能基础设施和 AI 算法包，包括标准算法接入、运行环境托管、线上线下资源对接等服务，帮助企业快速低成本的构建专属智能应用。
不论以领路者身份做东办赛，还是在后方提供基础设施，都符合阿里一贯的平台思维和生态打法。也许是得益于这种开放思路，阿里的 AI 工程师总能比别人更早发现产业痛点，更懂得以需求为导向推动产品创新。
比如，AI 肺结节检测准确率逼近理论极值，但为什么医生们并不感冒？甚至有医院同时部署多个公司的多种算法，但并不特别依赖任何一种。
答案也不复杂，单项的肺结节检测技术，即便在真实场景中的表现不输于实验室，也无法提升影像科医生的整体效率。肺部疾病种类多样，肺结节只是其一，医生阅片不可能只排查这一种，机器不能真正减轻医生负担。
基于这样的判断，阿里团队在肺结节顶级赛事夺冠之后，马上投入了肺部综合诊断方案的研发，实现对六种常见肺部病变的影像诊断：肺密度增高影、肺索条、肺大泡、动脉硬化、淋巴结钙化和肺结节。综合方案涵盖了大部分肺部疾病的早期筛查，不论对于医院影像科还是体检机构，都有非常现实的作用。这项技术现已通过阿里云对外输出，累计服务近千万用户。
医学集成了人类这个物种最大的自负和自卑，我们一方面相信，这门凭借代际智慧和临床经验立身的专业充满不可言说的经验性，没有谁能比人做得更好；另一方面我们深知，我们对自己身体的了解非常有限，机器又如何能懂得更多？
没人怀疑AI医学影像符合未来医学的发展趋势，但研发者们仍要不断证明 AI 的价值，让医疗 AI 的服务模式跟上技术进步的节奏。
