How AI is making it easier to diagnose disease | Pratik Shah

86,447 views ・ 2018-08-21

TED


請雙擊下方英文字幕播放視頻。

譯者: Lilian Chiu 審譯者: Helen Chang
00:13
Computer algorithms today are performing incredible tasks
0
13280
3856
現今的電腦演算法能夠執行 很了不起的工作任務,
00:17
with high accuracies, at a massive scale, using human-like intelligence.
1
17160
4736
有高度的精確性,規模可以很大, 且用的是類似人類的智慧。
00:21
And this intelligence of computers is often referred to as AI
2
21920
3936
這種電腦的智慧通常被稱為 AI,
00:25
or artificial intelligence.
3
25880
1856
也就是人工智慧。
00:27
AI is poised to make an incredible impact on our lives in the future.
4
27760
4200
人工智慧已經準備好要對 我們未來的生活造成衝擊。
00:32
Today, however, we still face massive challenges
5
32880
3936
然而我們現今仍然面臨很大的挑戰,
00:36
in detecting and diagnosing several life-threatening illnesses,
6
36840
3496
包括偵測與診斷數種 會威脅生命的疾病,
00:40
such as infectious diseases and cancer.
7
40360
2360
比如感染性疾病以及癌症。
00:44
Thousands of patients every year
8
44000
2296
每年,有數千名病人
00:46
lose their lives due to liver and oral cancer.
9
46320
2800
因為肝癌或口腔癌而喪命。
00:49
Our best way to help these patients
10
49880
2696
若要幫助這些病人的最好方法
00:52
is to perform early detection and diagnoses of these diseases.
11
52600
4320
就是早期偵測並診斷出這些疾病。
00:57
So how do we detect these diseases today, and can artificial intelligence help?
12
57880
4160
現今我們要如何偵測出這些疾病? 人工智慧能幫得上忙嗎?
01:03
In patients who, unfortunately, are suspected of these diseases,
13
63920
3656
對於很不幸被懷疑可能 得了這些疾病的病人,
01:07
an expert physician first orders
14
67600
2656
專業的醫生首先會囑咐
01:10
very expensive medical imaging technologies
15
70280
2616
採用非常昂貴的醫療成像技術,
01:12
such as fluorescent imaging, CTs, MRIs, to be performed.
16
72920
4096
例如螢光成像、 電腦斷層掃瞄、核磁共振。
01:17
Once those images are collected,
17
77040
2296
一旦收集到了這些影像,
01:19
another expert physician then diagnoses those images and talks to the patient.
18
79360
4520
會有另一位專業醫生根據 這些影像做診斷,並和病人談。
01:24
As you can see, this is a very resource-intensive process,
19
84520
3456
不難看出,這是非常耗資源的過程,
01:28
requiring both expert physicians, expensive medical imaging technologies,
20
88000
4416
需要專業的醫生 和昂貴的醫療成像技術兩者,
01:32
and is not considered practical for the developing world.
21
92440
3096
而這在開發中國家是不實際的;
01:35
And in fact, in many industrialized nations, as well.
22
95560
3360
事實上,在許多工業化的國家亦然。
01:39
So, can we solve this problem using artificial intelligence?
23
99760
2880
所以,我們能用人工智慧 來解決這個問題嗎?
01:43
Today, if I were to use traditional artificial intelligence architectures
24
103840
4056
現今,若我要用傳統人工智慧架構
01:47
to solve this problem,
25
107920
1216
來解決這個問題,
01:49
I would require 10,000 --
26
109160
1456
我會需要一萬——
01:50
I repeat, on an order of 10,000 of these very expensive medical images
27
110640
4016
我重覆一次,大約一萬張 這種非常昂貴的醫療影像
01:54
first to be generated.
28
114680
1376
先被產生出來。
01:56
After that, I would then go to an expert physician,
29
116080
2896
產生出來後,接著去找專業醫生,
01:59
who would then analyze those images for me.
30
119000
2496
來為我分析這些影像。
02:01
And using those two pieces of information,
31
121520
2096
用這兩種資訊,
02:03
I can train a standard deep neural network or a deep learning network
32
123640
3656
我就能訓練標準的 深度類神經網路或深度學習網路
02:07
to provide patient's diagnosis.
33
127320
2136
來提供對病人的診斷。
02:09
Similar to the first approach,
34
129480
1736
和第一個方法很類似,
02:11
traditional artificial intelligence approaches
35
131240
2143
傳統人工智慧方法
02:13
suffer from the same problem.
36
133407
1449
也會遇到同樣的問題。
02:14
Large amounts of data, expert physicians and expert medical imaging technologies.
37
134880
4560
大量的資料、專業醫生, 以及專業醫療成像技術。
02:20
So, can we invent more scalable, effective
38
140320
4296
我們是否能發明 更有擴展性、更有效,
02:24
and more valuable artificial intelligence architectures
39
144640
3296
且更有價值的人工智慧架構,
02:27
to solve these very important problems facing us today?
40
147960
3056
來解決我們現今所面臨的 這些非常重要的問題?
02:31
And this is exactly what my group at MIT Media Lab does.
41
151040
3296
這就是我的團隊在麻省理工學院 媒體實驗室在做的事。
02:34
We have invented a variety of unorthodox AI architectures
42
154360
3856
我們已經發明了多種 非正統的人工智慧架構
02:38
to solve some of the most important challenges facing us today
43
158240
3176
來解決我們現今在醫療成像 及臨床實驗方面
02:41
in medical imaging and clinical trials.
44
161440
2200
所面臨的一些最重要的挑戰。
02:44
In the example I shared with you today, we had two goals.
45
164480
3056
在今天我和各位分享的 例子中,我們有兩個目標。
02:47
Our first goal was to reduce the number of images
46
167560
2976
我們的第一個目標是要減少
02:50
required to train artificial intelligence algorithms.
47
170560
3256
訓練人工智慧演算法 所需要的影像數量。
02:53
Our second goal -- we were more ambitious,
48
173840
2096
我們的第二個目標—— 我們的野心更大,
02:55
we wanted to reduce the use of expensive medical imaging technologies
49
175960
3736
我們想要減少使用昂貴醫療成像技術
02:59
to screen patients.
50
179720
1216
來篩選病人。
03:00
So how did we do it?
51
180960
1200
我們要怎麼做?
03:02
For our first goal,
52
182920
1216
針對第一個目標,
03:04
instead of starting with tens and thousands
53
184160
2056
不像傳統人工智慧一開始
03:06
of these very expensive medical images, like traditional AI,
54
186240
3016
要用到數萬張非常 昂貴的醫療影像,
03:09
we started with a single medical image.
55
189280
2056
我們反而從單一張醫療影像開始。
03:11
From this image, my team and I figured out a very clever way
56
191360
3776
從這張影像,我和我的團隊 想出了一個非常聰明的方法
03:15
to extract billions of information packets.
57
195160
2736
來取出數十億個資訊封包。
03:17
These information packets included colors, pixels, geometry
58
197920
3696
這些資訊封包包括用 顏色、像素、幾何學,
03:21
and rendering of the disease on the medical image.
59
201640
2536
在醫療影像上呈現疾病。
03:24
In a sense, we converted one image into billions of training data points,
60
204200
4336
在某種意義上,我們是把一張影像 轉變為數十億個訓練資料點,
03:28
massively reducing the amount of data needed for training.
61
208560
3536
大大減少了訓練所需要的資料量。
03:32
For our second goal,
62
212120
1216
至於第二個目標,
03:33
to reduce the use of expensive medical imaging technologies to screen patients,
63
213360
3856
也就是減少使用昂貴的 醫療成像技術來篩選病人,
03:37
we started with a standard, white light photograph,
64
217240
2856
我們一開始使用的是 一張病人的標準白光照片,
03:40
acquired either from a DSLR camera or a mobile phone, for the patient.
65
220120
4336
可以用數位單眼相機或手機來拍攝。
03:44
Then remember those billions of information packets?
66
224480
2456
接著,還記得 那數十億個資訊封包嗎?
03:46
We overlaid those from the medical image onto this image,
67
226960
3536
我們將那些來自醫療影像的 封包疊到這張影像上,
03:50
creating something that we call a composite image.
68
230520
2520
創造出我們所謂的合成影像。
03:53
Much to our surprise, we only required 50 --
69
233480
3296
很讓我們驚訝的是, 我們只需要五十張——
03:56
I repeat, only 50 --
70
236800
1336
我重覆一次,只要五十張——
03:58
of these composite images to train our algorithms to high efficiencies.
71
238160
3840
這種合成影像,就能把我們的 演算法訓練到很高效能的程度。
04:02
To summarize our approach,
72
242680
1336
總結一下我們的方法,
04:04
instead of using 10,000 very expensive medical images,
73
244040
3176
我們不需要使用一萬張 非常昂貴的醫療影像,
04:07
we can now train the AI algorithms in an unorthodox way,
74
247240
3016
我們現在可以用非正統的方法 來訓練人工智慧演算法,
04:10
using only 50 of these high-resolution, but standard photographs,
75
250280
4256
只要用五十張高解析度的 一般標準照片,
04:14
acquired from DSLR cameras and mobile phones,
76
254560
2496
用數位單眼相機或手機來拍攝即可,
04:17
and provide diagnosis.
77
257080
1536
這樣就能提供出診斷結果。
04:18
More importantly,
78
258640
1216
更重要的是,
04:19
our algorithms can accept, in the future and even right now,
79
259880
3136
在未來,甚至在現在, 我們的演算法能接受
04:23
some very simple, white light photographs from the patient,
80
263040
2816
病人非常簡單的白光照片,
04:25
instead of expensive medical imaging technologies.
81
265880
2440
取代昂貴的醫療成像技術。
04:29
I believe that we are poised to enter an era
82
269120
3096
我相信我們已經準備好 要進入一個新時代,
04:32
where artificial intelligence
83
272240
1936
在這個時代,人工智慧
04:34
is going to make an incredible impact on our future.
84
274200
2536
將會對我們的未來有很大的衝擊。
04:36
And I think that thinking about traditional AI,
85
276760
2456
想想傳統人工智慧,
04:39
which is data-rich but application-poor,
86
279240
2776
它在資料上很豐富, 但在應用上很有限,
04:42
we should also continue thinking
87
282040
1536
我們應該要持續思考
04:43
about unorthodox artificial intelligence architectures,
88
283600
3016
有沒有其他非正統的 人工智慧架構,
04:46
which can accept small amounts of data
89
286640
1936
能夠接受更少量的資料,
04:48
and solve some of the most important problems facing us today,
90
288600
2936
並解決一些現今我們 面臨最重要的問題,
04:51
especially in health care.
91
291560
1256
特別是健康照護問題。
04:52
Thank you very much.
92
292840
1216
非常謝謝。
04:54
(Applause)
93
294080
3840
(掌聲)
關於本網站

本網站將向您介紹對學習英語有用的 YouTube 視頻。 您將看到來自世界各地的一流教師教授的英語課程。 雙擊每個視頻頁面上顯示的英文字幕,從那裡播放視頻。 字幕與視頻播放同步滾動。 如果您有任何意見或要求,請使用此聯繫表與我們聯繫。

https://forms.gle/WvT1wiN1qDtmnspy7