The reporter | GuLei

Ai is a big arena these days, and it’s important to know how the giants play.

Among many tracks in this arena, AI Medical has aroused endless imagination and frequent layout of giants with its huge potential market. In October 2016, Baidu released “Baidu Medical Brain”. In March 2017, Ali Cloud released “ET Medical Brain”; In August 2017, Tencent officially joined the battle of AI medical with the debut of its first AI medical imaging product “Tencent Miying”.

In the above layout, except Tencent releases a single medical image product, Baidu and Ali release ecological platforms, making people feel slightly abstract. How do they operate? Where are we now? Where do they focus on the AI health track? These are topics that startups and investors care about all the time.

Recently, in the “Cloud computing experts into Jingyi offline activities” held by Ali Cloud computing community, AI Technology Base camp and Ali cloud ET medical brain head Keegan had an exclusive dialogue, in order to enable readers to get rid of the current development of AI medical fog, see how Ali is playing on this popular track.

Li Zhaorong, name: Keegan, head of ali Cloud ET medical brain product operation, promoting artificial intelligence and data intelligence capabilities to serve the medical industry, once led the product design work of Ali Health Diegu Cloud HIS platform, network hospital, Ali Cloud ET medical brain, etc. Graduated from the University of Hong Kong with a Bachelor of Science in Biology with a minor in Economics; Master of Economics (Research), London School of Economics and Political Science (LSE).

Alet medical Brain is positioned as a medical operating system, providing infrastructure and standardized services for developers

In March 2017, ali cloud “ET medical brain” in shenzhen the cloud conference debut, xiao-ming hu President ali cloud to introduce the following: “the deep learning technology, ET training through vast amounts of data as an example to machine on a particular task, namely by studying cases of data by computer to improve the skill. At present, ET has a number of medical capabilities, and hopes to play the role of doctor’s assistant in the fields of patient virtual assistant, medical imaging, precision medicine, efficacy discovery, new drug research and development, health management and so on.”

Keegan went on to explain that ET Brain now plays differently than it did when it was launched in March. Ali has developed some medical algorithms before, but the focus of the current work is not to develop by itself, but to use the computing resources and business resources of Ali Cloud to help the developers of algorithms, so that their algorithms can be better used by customers or hospitals.

You can also understand this model as medical brain is like a medical operating system on which other algorithm teams build their own apps. In addition to some security management, operation and maintenance support already provided, Ali will gradually make the operating system thicker. In the future, NO matter a whole set of services or just an algorithm package, ET medical Brain will use some standardized ways to make them operate in hospitals.

There are two reasons for this shift, According to Keegan:

  1. Medical algorithms are more specific than general-purpose algorithms such as speech recognition. For example, the algorithms and tuning processes involved in pulmonary nodule detection and sugar mesh screening are relatively independent. And the application of artificial intelligence combined with the medical field is just beginning, compared with the rich medical scene, less than 1%.

  2. In the process of successful commercialization of ARTIFICIAL intelligence technology, Keegan believes that the importance of algorithms only accounts for part of it. The rest is more important in the process of integration of algorithms into hospital business systems. At the same time, to mobilize the wisdom of the whole industry to develop algorithms, Ali in the rear to provide integration and infrastructure support is more in line with ali Cloud’s overall ecological approach.

What can Alibaba offer companies developing algorithms?

What are the mutual interests of Ali and the algorithm companies developing on top of the medical Brain “operating system”? In response, Keegan sees Alibaba as offering developers or partners two things:

  • Algorithm development stage: it can provide developers with the lease of computing resources and the call of some algorithm packages.

In 2015, AliYun released the Platform of Artificial Intelligence (PAI), the first machine learning Platform in China, which not only can realize high-performance cloud computing to reduce storage and computing costs, but also has the corresponding tools and algorithm library to lower the technical threshold. In March 2017, AliYun upgraded PAI 2.0, which is now compatible with many deep learning frameworks and optimizes video memory, communication, and hardware/software collaboration.

  • Business cooperation between developers and hospitals: the implementation of business can be more efficient.

Keegan gives an example of a company that makes CDSS (Clinical Decision Sopport System). If they already have an algorithm model deployed to a hospital, they can only put a few machines in the hospital without using Alicloud. When algorithm optimization, log retrieval, troubleshooting, and system upgrade are required, engineers are required to perform offline maintenance.

Alibaba Cloud itself has data connection with many hospitals. If these companies put algorithms in the cloud, they can provide services to hospitals through dedicated lines. In this way, daily operations can be carried out in the cloud, and even a small company can provide technical support to hundreds of hospitals.

What does it take for an algorithm company?

Keegan told AI Tech Base that because Ali already has a lot of hospital resources, it often acts as a middleman in addition to providing infrastructure. So far, Alibaba has not charged any handling fees, but it does charge basic resources for data storage and transmission.

Position of medical team in Ali AI system: front desk in Zhongtai

Many people familiar with Ali know that at the end of 2015, Alibaba Group announced a comprehensive adjustment to its organizational structure, aiming to build a powerful zhongtai integrating The technology and data capabilities of Ali products, so as to form an organizational and business system of “big Zhongtai, small front desk”, so as to make front-line business more flexible and agile.

Keegan explained to AI technology base that the so-called middle platform is the technical capability provider, which is mainly from the internal iDST (Institute of Data Science&Technologies), AI LABS, the customer service team and external partners. For example, yitu Technology has entered the medical brain and other companies.

The medical brain is at the front of a center stage, keegan says, where a lot of scientists are working on a single point of capability that may not constitute the screening of a pulmonary nodule. For example, he may be doing a moving image recognition, or a still image recognition. Medical Brain will package his research into a concept with a business scenario before he can take it out.

How to package an algorithm into a product is a very complicated thing. According to Keegan, it can sometimes take a month to get an algorithm ready for market use, but it can take two to three months for the package to become a product, with an API-like process to integrate security, billing, single sign-on and other productized features. In the process of delivery to the hospital, some technical work, such as dynamic deployment, SDK and local image, may also be done to help the algorithm to better land in the hospital.

The front desk category of Ali for medical scenarios will include Alipay, Future Hospital, Ali Health, Dingding and other products.

ET brain structure and operation

Keegan told AI Technology Base that in addition to ET medical brain, Ali Cloud also has ET industrial brain, city brain, e-commerce brain, etc. These brains are parallel structure relationship, but sometimes there will be some cooperation, they all belong to Ali Cloud Flying.

But it was more than that, keegan added, it will come to big China architecture of the ali, we and our technical support team cooperation is very organic, for example, we have to do a very important thing, lack of front resources, may I go to China allocate, so is very flexible in terms of the number, sometimes more there will be hundreds of people in the medical side.

This mid-stage configuration has its benefits, Keegan explains. Many companies come up with a new medical division with 50 employees. In fact, it’s dangerous to have so many fixed resources at a time when the AI health space is changing so fast and no one has yet found a point where they can actually get the money, feed the team and generate a business.

Which hospitals are currently cooperating with? The in-depth cooperation is mainly in Zhejiang and Beijing

As for which hospitals have established and cooperated with at present, Keegan said that there have been many cooperation with hospitals, including Zhejiang No.1, Zhejiang No.2, Shanghai Huashan Hospital and Xinhua Hospital, but the depth of cooperation will be different. Generally speaking, the in-depth cooperation is concentrated in Zhejiang and Beijing area.

In terms of collaboration mode, Keegan said that the current binding mode with each collaboration site is different, but in general, the medical brain can provide resources for hospitals at two levels:

  • Infrastructure like computing resources

  • Custom algorithm development services

In the process of landing, what difficulties did you encounter?

1. Laboratory results cannot be reproduced in actual scenarios

When asked if there were any difficulties in getting the business off the ground, Keegan admitted that a very common but very important problem was that the data from the lab did not match the actual results from the hospital.

Screening for pulmonary nodules is very popular in the field of robotic screening because of the large number of publicly available data sets for pulmonary nodules. Public datasets have two characteristics:

  • One is that the wafers are very fine, often 1 mm or 1.5 mm thick.

  • One is that the integrity of the film is high, and there is rarely a lack of film.

The accuracy of machine reading is high on these two premises.

But in fact, in terms of CT layer thickness, the incision in the hospital is usually 5 mm, while the physical examination center is at least 5 mm, and there may be 1 cm thickness; In addition, the machine of hospital and physical examination center will often be short of film, and the missing is not one or two, which greatly affects the accuracy of machine reading film.

2. Data impassability in the hospital: it is difficult to get through and obtain the data effectively in the hospital

When asked about the problem of data silos between hospitals, Keegan laughs: “It’s not a major problem at the moment. The problem is that data is not even available within hospitals.”

Why is this a problem? Many people might say, isn’t it enough to integrate the imaging data in the hospital with the corresponding patient’s historical electronic medical record and learn from it? In practice, keegan explains, there was one big difficulty — getting the data.

This is because historical Electronic Medical records tend to have multiple EMR companies, and imaging data tends to be Packs or on-premise, or even in a third-party cloud, with zero interaction between the companies. It’s very difficult for most of the current medical imaging companies to deal with EMR or Packs, so it’s very difficult to get access to internal hospital data effectively, even without mechanisms and processes.

In the United States, there’s a HIPAA (Health Insurance Privacy and Accountability Act), which is very strict: Patients in the hospital took a film, the film is owned by many common data, including hospitals, patients, doctors, machines and software, if you want to be another person to use, the need to have the ownership of the data, the parties shall jointly sign can, can get a movie to see how hard it is data.

However, There is no such mechanism in China at present, which may lead to excessive development due to misuse of data on the one hand. On the other hand, there is no mature mechanism and process to go through, and people may also be told that no data can be used, resulting in excessive protection. There is not yet a good balance between these areas, so companies working on them are struggling.

Keegan believes that these problems will not change in the short term, but will only be solved slowly and cooperatively, how to face the data and produce valuable applications without causing data leakage.

Progress of Tianchi Medical AI Competition

It is worth mentioning that at the Cloud Computing Conference in Shenzhen in March this year, along with the release of ET Medical Brain, Ali also announced the launch of Tianchi Medical AI series competition, aiming to find the intelligent judgment optimal algorithm for early lung cancer diagnosis, so that the machine can assist doctors with diagnosis through original CT images.

According to AI Science and Technology Base camp, the number of participating teams has reached 2,887 by the time of writing, and the second round has ended, and the final defense will be held on October 11-12. According to Keegan, the competition is going relatively well so far, and Ali has selected a few good teams from the contestants to use their abilities for medical brain.

Meanwhile, Ali is now preparing for the second phase of the competition, which is open for recruitment…

AI medical breakthrough and Landing: Landing doesn’t sound sexy

When it comes to robotic screening for cervical cancer, lung and thyroid nodules, which are now in the press, Keegan says these applications seem to be more popular at the grassroots level.

Why is this the result? Since the training data used by the machine was manually annotated, no one could claim that the machine was already better than all the doctors. The good doctors thought they were better than the AI. At the same time, take the reading of pulmonary nodules as an example, the doctor does not only see pulmonary nodules in a film, but he has to see them again anyway, so the power of the machine is not strong. This is where AI’s current medical awkwardness lies.

As for the combination of AI and medical care, Keegan said he is more interested in the details of many medical procedures in hospitals, which can be optimized by machine learning and algorithms at any time.

For example, a patient swipes his medical insurance card to book a CT scan, and the system will recommend him an examination time. During this process, will there be some algorithms or scheduling models in operation, which will integrate some information of the patient — local patient or non-local patient? Medicare patients or unmedicare patients? Is it a SINGLE CT or is it waiting for surgery tomorrow? Emergency surgery or elective surgery? Is the blood ready for surgery?

These are happening now, and they are producing as much value as those machines, but they are getting less attention, Laments Mr Keegan.

Technical breakthrough or business breakthrough?

As for the definition of breakthrough, Keegan believes that breakthrough is divided into technical breakthrough and business breakthrough. For example, the accuracy of speech recognition is from 60% to 90%, which is a huge breakthrough in technology, but zero breakthrough in business, because speech recognition accuracy is less than 95%, it is not available. Going from 94 percent to 96 percent accuracy is a huge business breakthrough, and we already have a lot of meetings that use voice recognition directly as captions or shorthand.

As for current machine video readers, Keegan thinks they can only be used in certain scenarios. For example, the early screening technology for cervical cancer is developed by a company called Wuhan Landin, and the application scenario of their equipment is clearly targeted at grassroots hospitals.

In the whole AI medical ecology, what does Ali focus on?

So far, Ali’s focus is not on the news that a particular algorithm has landed in a particular hospital. Keegan believes that these are not replicable.

Ali’s focus is on the process of thinking about how to integrate the model with the real world of a hospital, from an AI algorithm to a usable software.

So, the goal of the medical brain is to help more good applications be implemented in hospitals, and we’re ready to do that for years, Keegan concluded.