ABOUT THE SPEAKER
Pratik Shah - Medical technologist
Dr. Pratik Shah creates novel intersections between engineering, medical imaging, machine learning and medicine.

Why you should listen

Dr. Shah's research program at the MIT Media Lab develops scalable and low-cost diagnostics and therapeutics. His ongoing research areas at MIT include: 1) artificial intelligence and machine learning methods for detection of cancer biomarkers using standard photographs vs. expensive medical images; 2) unorthodox artificial intelligence and machine learning algorithms to design optimal and faster clinical trials and to reduce adverse effects on patients; and 3) low-cost and open source imaging devices, paper diagnostics, algorithms and mobile phones to improve public health and generate real-world data.

Clinical studies with Pratik's medical technologies have revealed "missing sick" patients, who otherwise remain undiagnosed in conventional healthcare settings. Dr. Shah's graduate and postdoctoral research contributed to the discovery of a vaccine component to prevent pneumococcal (Streptococcus pneumoniae) diseases; the identification of new pathways, technologies and metabolites as antimicrobials to target gastrointestinal infections; and a nonprofit to deploy a low-cost water quality test for the developing world.

Past recognition for Dr. Shah includes the American Society for Microbiology's Raymond W. Sarber national award, the Harvard Medical School and Massachusetts General Hospitals ECOR Fund for Medical Discovery postdoctoral fellowship, the AAAS-Lemelson Invention Ambassador Award and a TED Fellowship. Pratik has been an invited discussion leader at Gordon Research Seminars; a speaker at Cold Spring Harbor Laboratories, Gordon Research Conferences and IEEE bioengineering conferences; and a peer reviewer for leading scientific publications and funding agencies. Pratik has a BS, MS, and a PhD in Microbiology and completed fellowship training at The Broad Institute of MIT and Harvard, Massachusetts General Hospital and Harvard Medical School.

More profile about the speaker
Pratik Shah | Speaker | TED.com
TEDGlobal 2017

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

派屈克 · 沙: 人工智慧如何讓診斷疾病變得容易

Filmed:
1,571,835 views

現今的人工智慧演算法需要數萬張昂貴的醫療影像才能偵測出病人的疾病。如果我們能大大減少訓練人工智慧所需要的資料量,讓診斷變得更低成本且高效益呢?TED 成員派屈克 · 沙已經開發出一項技術,只需五十張影像就可以發展出有用的演算法,甚至可用醫生手機拍的照片,就能提供診斷結果。來聽聽這場演說,進一步了解這個分析醫療資訊的新方法如何能協助及早偵測出威脅生命的疾病,並將人工智慧協助的診斷方法帶到世上更多的健康照護場所。
- Medical technologist
Dr. Pratik Shah creates novel intersections between engineering, medical imaging, machine learning and medicine. Full bio

Double-click the English transcript below to play the video.

00:13
Computer電腦 algorithms算法 today今天
are performing執行 incredible難以置信 tasks任務
0
1280
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精度, at a massive大規模的 scale規模,
using運用 human-like類人 intelligence情報.
1
5160
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情報 of computers電腦
is often經常 referred簡稱 to as AIAI
2
9920
3936
這種電腦的智慧通常被稱為 AI,
00:25
or artificial人造 intelligence情報.
3
13880
1856
也就是人工智慧。
00:27
AIAI is poised準備 to make an incredible難以置信 impact碰撞
on our lives生活 in the future未來.
4
15760
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今天, however然而,
we still face面對 massive大規模的 challenges挑戰
5
20880
3936
然而我們現今仍然面臨很大的挑戰,
00:36
in detecting檢測 and diagnosing診斷
several一些 life-threatening危及生命 illnesses疾病,
6
24840
3496
包括偵測與診斷數種
會威脅生命的疾病,
00:40
such這樣 as infectious傳染病 diseases疾病 and cancer癌症.
7
28360
2360
比如感染性疾病以及癌症。
00:44
Thousands成千上萬 of patients耐心 every一切 year
8
32000
2296
每年,有數千名病人
00:46
lose失去 their lives生活
due應有 to liver and oral口服 cancer癌症.
9
34320
2800
因為肝癌或口腔癌而喪命。
00:49
Our best最好 way to help these patients耐心
10
37880
2696
若要幫助這些病人的最好方法
00:52
is to perform演出 early detection發現
and diagnoses診斷 of these diseases疾病.
11
40600
4320
就是早期偵測並診斷出這些疾病。
00:57
So how do we detect檢測 these diseases疾病 today今天,
and can artificial人造 intelligence情報 help?
12
45880
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耐心 who, unfortunately不幸,
are suspected嫌疑 of these diseases疾病,
13
51920
3656
對於很不幸被懷疑可能
得了這些疾病的病人,
01:07
an expert專家 physician醫師 first orders命令
14
55600
2656
專業的醫生首先會囑咐
01:10
very expensive昂貴
medical imaging成像 technologies技術
15
58280
2616
採用非常昂貴的醫療成像技術,
01:12
such這樣 as fluorescent imaging成像,
CTsCts, MRIs核磁共振成像, to be performed執行.
16
60920
4096
例如螢光成像、
電腦斷層掃瞄、核磁共振。
01:17
Once一旦 those images圖片 are collected,
17
65040
2296
一旦收集到了這些影像,
01:19
another另一個 expert專家 physician醫師 then diagnoses診斷
those images圖片 and talks會談 to the patient患者.
18
67360
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資源密集型 process處理,
19
72520
3456
不難看出,這是非常耗資源的過程,
01:28
requiring要求 both expert專家 physicians醫師,
expensive昂貴 medical imaging成像 technologies技術,
20
76000
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考慮 practical實際的
for the developing發展 world世界.
21
80440
3096
而這在開發中國家是不實際的;
01:35
And in fact事實, in many許多
industrialized工業化 nations國家, as well.
22
83560
3360
事實上,在許多工業化的國家亦然。
01:39
So, can we solve解決 this problem問題
using運用 artificial人造 intelligence情報?
23
87760
2880
所以,我們能用人工智慧
來解決這個問題嗎?
01:43
Today今天, if I were to use traditional傳統
artificial人造 intelligence情報 architectures架構
24
91840
4056
現今,若我要用傳統人工智慧架構
01:47
to solve解決 this problem問題,
25
95920
1216
來解決這個問題,
01:49
I would require要求 10,000 --
26
97160
1456
我會需要一萬——
01:50
I repeat重複, on an order訂購 of 10,000
of these very expensive昂貴 medical images圖片
27
98640
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產生.
28
102680
1376
先被產生出來。
01:56
After that, I would then go
to an expert專家 physician醫師,
29
104080
2896
產生出來後,接著去找專業醫生,
01:59
who would then analyze分析
those images圖片 for me.
30
107000
2496
來為我分析這些影像。
02:01
And using運用 those two pieces of information信息,
31
109520
2096
用這兩種資訊,
02:03
I can train培養 a standard標準 deep neural神經 network網絡
or a deep learning學習 network網絡
32
111640
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提供 patient's耐心 diagnosis診斷.
33
115320
2136
來提供對病人的診斷。
02:09
Similar類似 to the first approach途徑,
34
117480
1736
和第一個方法很類似,
02:11
traditional傳統 artificial人造
intelligence情報 approaches方法
35
119240
2143
傳統人工智慧方法
02:13
suffer遭受 from the same相同 problem問題.
36
121407
1449
也會遇到同樣的問題。
02:14
Large amounts of data數據, expert專家 physicians醫師
and expert專家 medical imaging成像 technologies技術.
37
122880
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發明 more scalable可擴展性, effective有效
38
128320
4296
我們是否能發明
更有擴展性、更有效,
02:24
and more valuable有價值 artificial人造
intelligence情報 architectures架構
39
132640
3296
且更有價值的人工智慧架構,
02:27
to solve解決 these very important重要
problems問題 facing面對 us today今天?
40
135960
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究竟
what my group at MITMIT Media媒體 Lab實驗室 does.
41
139040
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發明 a variety品種
of unorthodox非正統的 AIAI architectures架構
42
142360
3856
我們已經發明了多種
非正統的人工智慧架構
02:38
to solve解決 some of the most important重要
challenges挑戰 facing面對 us today今天
43
146240
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成像 and clinical臨床 trials試驗.
44
149440
2200
所面臨的一些最重要的挑戰。
02:44
In the example I shared共享
with you today今天, we had two goals目標.
45
152480
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目標 was to reduce減少
the number of images圖片
46
155560
2976
我們的第一個目標是要減少
02:50
required需要 to train培養
artificial人造 intelligence情報 algorithms算法.
47
158560
3256
訓練人工智慧演算法
所需要的影像數量。
02:53
Our second第二 goal目標 -- we were more ambitious有雄心,
48
161840
2096
我們的第二個目標——
我們的野心更大,
02:55
we wanted to reduce減少 the use
of expensive昂貴 medical imaging成像 technologies技術
49
163960
3736
我們想要減少使用昂貴醫療成像技術
02:59
to screen屏幕 patients耐心.
50
167720
1216
來篩選病人。
03:00
So how did we do it?
51
168960
1200
我們要怎麼做?
03:02
For our first goal目標,
52
170920
1216
針對第一個目標,
03:04
instead代替 of starting開始
with tens and thousands數千
53
172160
2056
不像傳統人工智慧一開始
03:06
of these very expensive昂貴 medical images圖片,
like traditional傳統 AIAI,
54
174240
3016
要用到數萬張非常
昂貴的醫療影像,
03:09
we started開始 with a single medical image圖片.
55
177280
2056
我們反而從單一張醫療影像開始。
03:11
From this image圖片, my team球隊 and I
figured想通 out a very clever聰明 way
56
179360
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提取 billions數十億
of information信息 packets.
57
183160
2736
來取出數十億個資訊封包。
03:17
These information信息 packets
included包括 colors顏色, pixels像素, geometry幾何
58
185920
3696
這些資訊封包包括用
顏色、像素、幾何學,
03:21
and rendering翻譯 of the disease疾病
on the medical image圖片.
59
189640
2536
在醫療影像上呈現疾病。
03:24
In a sense, we converted轉換 one image圖片
into billions數十億 of training訓練 data數據 points,
60
192200
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大規模 reducing減少 the amount of data數據
needed需要 for training訓練.
61
196560
3536
大大減少了訓練所需要的資料量。
03:32
For our second第二 goal目標,
62
200120
1216
至於第二個目標,
03:33
to reduce減少 the use of expensive昂貴 medical
imaging成像 technologies技術 to screen屏幕 patients耐心,
63
201360
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開始 with a standard標準,
white白色 light photograph照片,
64
205240
2856
我們一開始使用的是
一張病人的標準白光照片,
03:40
acquired後天 either from a DSLRDSLR camera相機
or a mobile移動 phone電話, for the patient患者.
65
208120
4336
可以用數位單眼相機或手機來拍攝。
03:44
Then remember記得 those
billions數十億 of information信息 packets?
66
212480
2456
接著,還記得
那數十億個資訊封包嗎?
03:46
We overlaid覆蓋 those from
the medical image圖片 onto this image圖片,
67
214960
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創建 something
that we call a composite綜合 image圖片.
68
218520
2520
創造出我們所謂的合成影像。
03:53
Much to our surprise,
we only required需要 50 --
69
221480
3296
很讓我們驚訝的是,
我們只需要五十張——
03:56
I repeat重複, only 50 --
70
224800
1336
我重覆一次,只要五十張——
03:58
of these composite綜合 images圖片 to train培養
our algorithms算法 to high efficiencies效率.
71
226160
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總結 our approach途徑,
72
230680
1336
總結一下我們的方法,
04:04
instead代替 of using運用 10,000
very expensive昂貴 medical images圖片,
73
232040
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培養 the AIAI algorithms算法
in an unorthodox非正統的 way,
74
235240
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運用 only 50 of these high-resolution高分辨率,
but standard標準 photographs照片,
75
238280
4256
只要用五十張高解析度的
一般標準照片,
04:14
acquired後天 from DSLRDSLR cameras相機
and mobile移動 phones手機,
76
242560
2496
用數位單眼相機或手機來拍攝即可,
04:17
and provide提供 diagnosis診斷.
77
245080
1536
這樣就能提供出診斷結果。
04:18
More importantly重要的,
78
246640
1216
更重要的是,
04:19
our algorithms算法 can accept接受,
in the future未來 and even right now,
79
247880
3136
在未來,甚至在現在,
我們的演算法能接受
04:23
some very simple簡單, white白色 light
photographs照片 from the patient患者,
80
251040
2816
病人非常簡單的白光照片,
04:25
instead代替 of expensive昂貴
medical imaging成像 technologies技術.
81
253880
2440
取代昂貴的醫療成像技術。
04:29
I believe that we are poised準備
to enter輸入 an era時代
82
257120
3096
我相信我們已經準備好
要進入一個新時代,
04:32
where artificial人造 intelligence情報
83
260240
1936
在這個時代,人工智慧
04:34
is going to make an incredible難以置信
impact碰撞 on our future未來.
84
262200
2536
將會對我們的未來有很大的衝擊。
04:36
And I think that thinking思維
about traditional傳統 AIAI,
85
264760
2456
想想傳統人工智慧,
04:39
which哪一個 is data-rich數據豐富 but application-poor應用程式窮,
86
267240
2776
它在資料上很豐富,
但在應用上很有限,
04:42
we should also continue繼續 thinking思維
87
270040
1536
我們應該要持續思考
04:43
about unorthodox非正統的 artificial人造
intelligence情報 architectures架構,
88
271600
3016
有沒有其他非正統的
人工智慧架構,
04:46
which哪一個 can accept接受 small amounts of data數據
89
274640
1936
能夠接受更少量的資料,
04:48
and solve解決 some of the most important重要
problems問題 facing面對 us today今天,
90
276600
2936
並解決一些現今我們
面臨最重要的問題,
特別是健康照護問題。
04:51
especially特別 in health健康 care關心.
91
279560
1256
04:52
Thank you very much.
92
280840
1216
非常謝謝。
04:54
(Applause掌聲)
93
282080
3840
(掌聲)
Translated by Lilian Chiu
Reviewed by Helen Chang

▲Back to top

ABOUT THE SPEAKER
Pratik Shah - Medical technologist
Dr. Pratik Shah creates novel intersections between engineering, medical imaging, machine learning and medicine.

Why you should listen

Dr. Shah's research program at the MIT Media Lab develops scalable and low-cost diagnostics and therapeutics. His ongoing research areas at MIT include: 1) artificial intelligence and machine learning methods for detection of cancer biomarkers using standard photographs vs. expensive medical images; 2) unorthodox artificial intelligence and machine learning algorithms to design optimal and faster clinical trials and to reduce adverse effects on patients; and 3) low-cost and open source imaging devices, paper diagnostics, algorithms and mobile phones to improve public health and generate real-world data.

Clinical studies with Pratik's medical technologies have revealed "missing sick" patients, who otherwise remain undiagnosed in conventional healthcare settings. Dr. Shah's graduate and postdoctoral research contributed to the discovery of a vaccine component to prevent pneumococcal (Streptococcus pneumoniae) diseases; the identification of new pathways, technologies and metabolites as antimicrobials to target gastrointestinal infections; and a nonprofit to deploy a low-cost water quality test for the developing world.

Past recognition for Dr. Shah includes the American Society for Microbiology's Raymond W. Sarber national award, the Harvard Medical School and Massachusetts General Hospitals ECOR Fund for Medical Discovery postdoctoral fellowship, the AAAS-Lemelson Invention Ambassador Award and a TED Fellowship. Pratik has been an invited discussion leader at Gordon Research Seminars; a speaker at Cold Spring Harbor Laboratories, Gordon Research Conferences and IEEE bioengineering conferences; and a peer reviewer for leading scientific publications and funding agencies. Pratik has a BS, MS, and a PhD in Microbiology and completed fellowship training at The Broad Institute of MIT and Harvard, Massachusetts General Hospital and Harvard Medical School.

More profile about the speaker
Pratik Shah | Speaker | TED.com