ADM-201 dump PMP dumps pdf SSCP exam materials CBAP exam sample questions

TensorFlow在医药方面的应用案例分析-视网膜成像 – 译学馆
未登陆,请登陆后再发表信息
最新评论 (0)
播放视频

TensorFlow在医药方面的应用案例分析-视网膜成像

Case Study: TensorFlow in Medicine - Retinal Imaging (TensorFlow Dev Summit 2017)

[播放音乐]
[MUSIC PLAYING]
LILY PENG: 大家好
LILY PENG: Hi, everyone.
我叫Lily
So, I’m Lily.
我在谷歌大脑部门的医疗成像小组工作
I work on our medical imaging team in Brain.
在此以前 我是一名医生
In a previous life, I was a doctor,
现在我已经转行成为Google的一名产品经理
and I’ve been repurposed as a product manager at Google.
[笑声]
[LAUGHTER]
我们组一直都在进行中的一个项目是
One of the projects that we’ve been working on in our group
利用深度学习来做视网膜成像
is using deep learning for retinal imaging.
特别是 我们正在关注一种
In particular, we are looking at a disease
被称为糖尿病视网膜病变(DR)的疾病
called diabetic retinopathy.
这个病除了很拗口 它其实
Other than a mouthful, it’s actually also
也是全世界增长最快的导致失明的原因
the fastest growing cause of blindness in the world,
由于它是糖尿病的一种并发症
and it’s because this is a complication of diabetes.
全世界有4.15亿人罹患糖尿病
There are 415 million people in the world with diabetes,
正是由于DR 或者说糖尿病视网膜病变
and each one of them is at risk for going
他们中的每一个都处于失明的危险中
blind due to what we call DR, or Diabetic Retinopathy.
防止失明的关键是常规检查
The key to preventing blindness is regular screening.
全世界所有的指导方针都是建议每年
Every guideline worldwide recommends about once a year
做一次检查 因为这种病在达到
screening, and it’s because this is pretty asymptomatic
无法逆转的失明状态之前 基本上是无症状的
until you get to a point where there’s irreversible visionloss.
而到那个时候再来干预治疗就有点太晚了
And at that point, it’s a little too late to intervene.
这个检查需要用一种专门的相机
This is done by taking a picture using a specialized camera
通过你的眼球 拍摄眼睛底部
of the back of the eye through your pupil,
然后再由医生对这些图像评级
and then a doctor grades these images.
我们在图像中寻找一些小的出血和小斑点
We look for these little hemorrhages and little spots
然后我们对它们进行评级 从无疾病到末期
on the image, and we grade them on a five-class scale
一共分为五级 其中末期又被称为
from no disease to the end stage, which
增生的糖尿病视网膜病变
is sort of proliferative DR.
在世界上很多地方 包括
In many places in the world, including
我们这个故事的起源地-印度
in India where our story originated,
都没有足够的医生来完成评级这件工作
there are just simply not enough doctors to do this task.
印度缺少大概127,000位眼科医生
In India, there is a shortage of 127,000-some eye doctors,
正是由于这种短缺和其它系统的问题
and because of this and other systematic issues,
大约有一半的人在被诊断之前就已经处于
about half of people actually suffer vision loss before
丧失视力的痛苦中
they’re even diagnosed.
考虑到这是完全可以预防的疾病
For something that’s completely preventable,
这种情况可以说是无法接受的
this is sort of unacceptable.
这是一张正在排队等待检查的
Here is a picture of some of the people
人们的照片
who are waiting in line to get screened.
所以 即使你去到提供检查的地方
Even if you get to a place where there is screening,
也有很长的等待时间
there is a long wait.
和很长的周转时间 因此 很多人最终
There is long turnaround time, and so a lot of people
没能得到治疗
end up being lost to care.
另一个问题是 即使有医生
The other issue is that even when available,
他们之间的差异也出乎意料的大
doctors are surprisingly variable.
在这张图里 每一种颜色代表此种疾病
Here in this graph, each color represents a different class
的一个等级 每一行
of category of disease, and each row
代表一个病人的眼底图像
is a patient image of that fundus image,
而每一列代表一位眼科医生
and each column represents an ophthalmologist.
这些都是有美国医师牌照的眼科医生
These are US board-certified ophthalmologists,
在我们试图解决这个问题的时候
and we had given them the test set
我们给了他们一些作为测试的病例
when we were trying to attack this problem
来看看他们每一位的评级
to see what the grades were for each of them.
你可以看到 当无疾病的时候
And as you can see, when there’s no disease,
他们的意见相当一致
there is pretty good agreement.
只有一位医生认为有病
There’s one person who thinks otherwise,
但是其他所有人都达成了一致
but everyone is the consensus is there.
然后 当处于末期的时候
And then, of course, you look at the end stage
有增生的病变
when there’s proliferative disease,
他们的意见也非常一致
there’s good agreement there, too.
但是在这两种情况之间 医生的意见
But in between, there’s actually a lot
有很大不同 对于到底属于哪一种
of variability and disagreement about where
有很大异议 尽管大家
this should actually fit, even though there
都知道其参考标准
are pretty well-known guidelines,
这是因为 基本上
and it’s because human beings in general
对于从那张图能看到什么 人类是无法做到
just aren’t super great at being very precise about what
完全精确的
we see in that image.
另外 你可以看到用黑色方框强调出来的两行
And of course, you can see the two highlighted rows there in black.
这些图片包括了书上的每个级别 对不对?
These images got every grade in the book, right?
所以 取决于你看的哪个医生 对你的处理方式
So depending on who you saw, your management
将会有所不同
would be kind of different.
过会我们会再回到这个话题
A little bit more about that later.
我们认为可以改善的部分是 来训练一个模型吧
Where we thought we could help was let’s train a model.
所以 事实上我们创建了一个标记工具
And so we actually built a labeling tool.
我们从13万张图片开始
We started off with 130,000 images,
从那时到现在 我们已经多了很多图片
and we’ve gotten much more since then,
我们雇了一个军队的眼科医生来
and we hired an army of ophthalmologists
帮助标记
to help us label.
对于这些图片 我们从54个
And from our 54 ophthalmologists,
眼科医生那里得到了88万个诊断
we got 880,000 diagnoses for these images.
你可以从前一页看到
And you can see from the previous slide
为什么我们要这么做 因为有时候需要
why we did that, because sometimes it took up to seven
多达七次读数才能得出一致的诊断
reads to get something consistent.
在将这些数据清理好并且
Then what we did after we got this data cleaned
完成标记后 我们就使用非常值得信任的Inception网络
up and labeled, we used our trusty, dusty inception
此网络已经成功完成了很多图像识别任务
network that works for a lot of image recognition tasks,
从猫猫到狗狗 再到黑素瘤 现在又到DR
from cats to puppies to melanoma, and now DR.
我们训练它来检测这种五级预测
And we trained it to detect these five class predictions,
并且也让它预测一些对于临床医生很需要知道的
but we also asked it to predict housekeeping things
内部管理事务
that may be important for a clinician
– 是否这张图片的质量
to know– whether or not this image is
足以作出评级
of sufficient quality for grading,
这幅图片是左眼还是右眼
whether or not this is a left or right eye.
有时我们也会困惑
Sometimes we get confused.
还有视野范围
And also the field of view, which
即你正在看的是视网膜的哪部分
is like what part of the retina you’re actually seeing.
然后 我们为它创建了一个前端
And then we built a front end to this.
我会在这里做一个演示
I’m going to try a demo here.
这个被我戏称为烤面包机
This is literally what I call a toaster.
我们试着拖拽移动一些东西
We try to drag and drop something.
事实上我不知道该如何-如何移动指针
I don’t know actually how to– how do I move the cursor?
哦 好了
Oh, there we are.
我要打开浏览器
So I’m going to open up a web browser,
希望它能正常工作
and hopefully that works.
我要把一张图片拖拽过来
I’m going to drag one of our images over.
我看不到
I can’t see it.
哦 在这
Oh, there we go.
它正在分析
And it analyzing.
它应该更快一些 但是这个演示……
It should be faster, but the demo
哦 分析完成了
gods– oh, it’s cooperating.
现在我们可以告诉你这里
So here we are able to tell you that there’s
有增生的病变
proliferative disease here.
但是没有DMU
There is no what we call DMU, which
DMU是一种不同种类的DR
is a different type of DR. And we
我们可以说这是处于中等和严重
are saying that this is somewhere between
之间的某个程度 而且它也确实是处于
moderate and severe, and this is indeed something
中等和严重之间
between moderate and severe.
这个例子稍微展示了一下它对于每个具体病例是如何工作的
I kind of showed you how it works on a case-by-case basis,
那么对于许多病例一起 它又是如何工作的呢?
but then how does it work over a lot of images?
事实上 我们已经在“美国医学会杂志”(JAMA)
Well, here we actually published and shared
上发表和分享了我们的工作
how we did this work in the “Journal of the American
这是我们使用的其中一个测试
Medical Association,” and this is one of the tests
或者说确认组
or the validation sets that we use.
这个模型之前并没有被此组数据训练过或者测试过
The model was not trained or tested on this previously.
对这9963幅图片 我们预测了其是否患上
And out of 9,963 images, we predicted whether or not
所涉及的疾病
it had referrable disease.
Y轴表示灵敏度
The y-axis is sensitivity.
X轴表示(1-Specificity)(特异性)
The x-axis is 1 minus specificity.
我们的算法 如果你
And our algorithm and the two black dots,
看得见的话 就是两个黑点
if you can actually see it, is in black.
那就是我们的算法
That’s our algorithm.
那些小的彩色点是
And then the little colored dots are
拥有美国医师执照的眼科医生
US board-certified ophthalmologists.
图中靠左边表示好 你可以看到 从本质上来说
And to the left is good, and you can see that essentially we’re
我们的表现和大多数眼科医生非常接近
very close to most of the ophthalmologists in terms of performance.
事实上 如果你看一下我们的F-分数
And in fact, if you look at our F-score,
比较一下算法的F-分数和“中位数”的F-分数
and you compare the algorithm’s F-score to that of the median,
“中位数”指的是处于这群眼科医生正中间的那些医生
ophthalmologists were sort of in the middle of the pack.
我们决定在JAMA发表文章的一个原因是
One of the reasons we also decided to publish in JAMA
因为我们相信参与到医学群体中
was because we believe that engaging the medical community
对于将这些技术传播到那些最终实际运用它们的人当中
is really important to get these technologies out
是非常重要的
into the hands of the people who could actually use them.
而且这也确实广受好评
It was actually quite well-received.
你能看到很多真正的医生引用了我们的工作
You can see some quotes from real doctors about our work,
这让我们非常振奋
and so we’re really excited about that.
那么 TensorFlow如何帮助我们呢?
How did TensorFlow help us?
我认为 在我们工作中的每一步 它都在
Well, every step of the way, I think
帮助我们迅速地设计模型从而真正地启动各项工作
it helped us really start with quick prototyping.
它帮助我们启动了系统架构和预训练模型
We had started our architecture, pre-trained models,
而且事实上 它让我们能够试验神经网络的
and we actually were able to try out different variations
各个不同变量 从而发现-
of neuro-networks, and we actually found–
我的意思是 真正地发现
I mean, we literally found that Inception–
Inception B3在目前
B3 at this point–
是最优的
worked the best.
但是我们可以非常迅速地试验很多东西
But we could try out things very quickly.
我们也可以做预训练
And we also pre-trained.
事实上 我们在经典的ImageNet上做预训练
So we actually pre-trained on the classic image net,
发现它对性能表现有很大的提高
and we found there was a boost in performance there.
它也帮助我们对规模进行实验 比如
It also helped us to experiment at scale, so
GPU支持和快速训练
GPU support and fast training.
从而让我们可以运行各种不同的实验
This allows us to run all these different experiments,
不同的标记
different sort of labeling.
比如增加新的标记或者不同的标记
If we had new labels or different labels,
它可以帮助我们完成这些
it kind of helped us do that.
最后 我觉得非常重要的一点是
And finally, what I think is really important
它让我们小组可以再利用已付出的努力
is that it really allowed our team to reinvest the efforts,
所以最大的难题不再是机器学习(ML)和训练了
so the blocker was no longer in machine learning and in training.
那一直是 ……
That’s always been like, you know ……
那都是很难完成的工作 如果你看看我们所做的
That’s been hard to do, and if you look at what we did here,
在这里 我们其实应用的是非常简单易懂的ML技术
we actually applied very straightforward ML techniques here.
真正的神奇佐料其实是
What was the magic sauce was actually
找出正确的问题 得到数据
finding the right problem, getting the data,
对图片的判断达成一致意见
getting agreement about what was in the image,
然后我们能够使用这整套
and then we were able to use this package, these tools that
这些工具可以训练模型 而且模型
were able to allow us to train the models, that actually
的实际表现非常非常棒
performed really, really well.
并且 它也能够让我们小组
And that also, then, allows our team
专注在验证算法上
to focus on validating the algorithm
以及想出如何将其运用到健康系统中
and figure out ways to deploy it into health systems, which
这本身就是一个巨大的挑战
in itself is a huge challenge.
我们下一步做什么呢?
What’s next for us?
我们训练了一个模型
We train a model.
它工作得非常好
It works really well.
现在 我们需要真正用临床检验它
Now we need to actually clinically validate it.
我们已经和印度的两家医院 Aravind和Sankara
We’ve been working with two hospitals in India, Aravind
取得合作 他们正在对之前我们所讲的算法
and Sankara, and they’re running clinical trials
进行临床实验
of the algorithm as we speak.
事实上 Aravind的实验已经完成 他们得到了
Actually, Aravind’s finished, and they
基本相同的结果-即我们的算法
have found essentially the same results–
比他们那儿的眼科医师的平均水平
that we were slightly better than the average
要稍微好一点
of their ophthalmologists there.
所以 我们现在正在和两家公司合作 一家是Alphabet旗下的
And so what we’re doing is working with a fellow Alphabet
专注于生命科学的公司-Verily 另一家是
company, Verily, that’s a life science-focused company,
硬件制造商-Nikon
and a hardware maker called Nikon.
你也许没有听说过这家小公司
You may not have heard of that little company.
目前的状况是 算法已经
But the idea is now that the algorithm
工作得很好了 瓶颈在于硬件
works pretty well, the bottleneck becomes
因为我们需要一种专门的照相机
the hardware, because we need a specialized camera
来拍摄这些图片
to take these pictures.
所以 我们与硬件制造商合作
So we’re working with the hardware manufacturers
希望能找到办法配置好用而且
to essentially figure out ways to deploy lightweight hardware
轻便的硬件
that’s easy to use, et cetera.
退一步说 我学医的一个
Taking a step back, one of the main reasons
主要原因是我是医学哲学博士
I got into medicine was I was an MD-PhD,
所以 对于可以将这个科学突破从案台引入到临床
and so I really was very excited about bringing breakthrough
我感到非常兴奋
science from bench to bedside.
当你在实行那些(模型)训练时
And there’s a part of you, when you go through training,
作为一名博士 一部分的你会认为
and you’re a PhD, and you’re like, this is never
这绝不可能成功 因为
going to happen, because it’s just not
它不可能解决这些问题的
possible to solve these problems.
但是有了TensorFlow以及我们在这里做的所有工作
And with TensorFlow and all the work that’s been done here,
却使得它成为了可能
it’s actually possible to do that.
我们可以训练算法
It’s possible to train algorithms
使它确实能够帮助医生将诊治
that really can help physicians deliver care
带给那些最需要的人们
where people need it most.
[鼓掌]
[APPLAUSE]
[播放音乐]
[MUSIC PLAYING]

发表评论

译制信息
视频概述

在TensorFlow的帮助下,利用深度学习来分析眼底照片,从而正确诊断出糖尿病视网膜病变。

听录译者

收集自网络

翻译译者

豆子

审核员

审核团1024

视频来源

https://www.youtube.com/watch?v=oOeZ7IgEN4o

相关推荐