1
00:00:01,333 --> 00:00:05,150
Computer algorithms today
are performing incredible tasks

2
00:00:05,174 --> 00:00:09,682
with high accuracies, at a massive scale,
using human-like intelligence.

3
00:00:10,053 --> 00:00:13,863
And this intelligence of computers
is often referred to as AI

4
00:00:13,887 --> 00:00:15,736
or artificial intelligence.

5
00:00:15,760 --> 00:00:20,427
AI is poised to make an incredible impact
on our lives in the future.

6
00:00:20,911 --> 00:00:24,807
Today, however, we still
face massive challenges

7
00:00:24,831 --> 00:00:28,325
in detecting and diagnosing
several life-threatening illnesses,

8
00:00:28,349 --> 00:00:30,854
such as infectious diseases and cancer.

9
00:00:32,023 --> 00:00:34,277
Thousands of patients, every year,

10
00:00:34,301 --> 00:00:37,434
lose their lives due to
liver and oral cancer.

11
00:00:37,894 --> 00:00:40,577
Our best way to help these patients

12
00:00:40,601 --> 00:00:44,910
is to perform early detection
and diagnoses of these diseases.

13
00:00:45,871 --> 00:00:50,260
So how do we detect these diseases today
and can artificial intelligence help?

14
00:00:51,966 --> 00:00:55,592
In patients who, unfortunately,
are suspected of these diseases,

15
00:00:55,616 --> 00:00:58,243
an expert physician first orders

16
00:00:58,267 --> 00:01:00,893
very expensive medical
imaging technologies

17
00:01:00,917 --> 00:01:04,668
such as fluorescent imaging, CTs,
MRIs, to be performed.

18
00:01:05,046 --> 00:01:07,339
Once those images are collected,

19
00:01:07,363 --> 00:01:12,107
another expert physician then diagnoses
those images and talks to the patient.

20
00:01:12,522 --> 00:01:15,958
As you can see, this is a very
resource-intensive process,

21
00:01:15,982 --> 00:01:20,418
requiring both expert physicians,
expensive medical imaging technologies,

22
00:01:20,442 --> 00:01:23,521
and is not considered practical
for the developing world.

23
00:01:23,545 --> 00:01:27,037
And in fact, in many
industrialized nations, as well.

24
00:01:27,784 --> 00:01:31,069
So, can we solve this problem
using artificial intelligence?

25
00:01:31,887 --> 00:01:35,903
Today, if I were to use traditional
artificial intelligence architectures

26
00:01:35,927 --> 00:01:37,133
to solve this problem,

27
00:01:37,157 --> 00:01:38,736
I would require 10,000 --

28
00:01:38,760 --> 00:01:42,664
I repeat, on an order of 10,000
of these very expensive medical images

29
00:01:42,688 --> 00:01:44,085
first to be generated.

30
00:01:44,109 --> 00:01:46,974
After that, I would then go
to an expert physician,

31
00:01:46,998 --> 00:01:49,514
who would then analyze
those images for me.

32
00:01:49,538 --> 00:01:51,696
And using those two pieces of information,

33
00:01:51,720 --> 00:01:55,299
I can train a standard deep neural network
or a deep learning network,

34
00:01:55,323 --> 00:01:57,459
to provide patient's diagnosis.

35
00:01:57,483 --> 00:01:59,198
Similar to the first approach,

36
00:01:59,222 --> 00:02:01,388
traditional artificial
intelligence approaches

37
00:02:01,412 --> 00:02:02,856
suffer from the same problem.

38
00:02:02,880 --> 00:02:07,322
Large amounts of data, expert physicians,
and expert medical imaging technologies.

39
00:02:08,354 --> 00:02:12,601
So can we invent more scalable, effective,

40
00:02:12,625 --> 00:02:15,934
and more valuable
artificial intelligence architectures

41
00:02:15,958 --> 00:02:18,692
to solve these very important
problems facing us today?

42
00:02:19,034 --> 00:02:22,049
And this is exactly what my group
at MIT Media Lab does.

43
00:02:22,375 --> 00:02:26,200
We have invented a variety
of unorthodox AI architectures

44
00:02:26,224 --> 00:02:29,426
to solve some of the most important
challenges facing us today

45
00:02:29,450 --> 00:02:31,902
in medical imaging and clinical trials.

46
00:02:32,482 --> 00:02:35,537
In the example I shared
with you today, we had two goals.

47
00:02:35,561 --> 00:02:38,522
Our first goal was to reduce
the number of images

48
00:02:38,546 --> 00:02:41,815
required to train artificial
intelligence algorithms.

49
00:02:41,839 --> 00:02:43,926
Our second goal -- we were more ambitious,

50
00:02:43,950 --> 00:02:47,712
we wanted to reduce the use
of expensive medical imaging technologies

51
00:02:47,736 --> 00:02:48,934
to screen patients.

52
00:02:48,958 --> 00:02:50,292
So how did we do it?

53
00:02:50,966 --> 00:02:52,149
For our first goal,

54
00:02:52,173 --> 00:02:54,220
instead of starting
with tens and thousands

55
00:02:54,244 --> 00:02:57,253
of these very expensive medical images,
like traditional AI,

56
00:02:57,277 --> 00:02:59,355
we started with a single medical image.

57
00:02:59,379 --> 00:03:03,125
From this image, my team and I
figured out a very clever way

58
00:03:03,149 --> 00:03:05,895
to extract billions
of information packets.

59
00:03:05,919 --> 00:03:09,628
These information packets
included colors, pixels, geometry

60
00:03:09,652 --> 00:03:12,180
and rendering of the disease
on the medical image.

61
00:03:12,204 --> 00:03:16,537
In a sense, we converted one image
into billions of training data points,

62
00:03:16,561 --> 00:03:19,544
massively reducing the amount of data
needed for training.

63
00:03:20,148 --> 00:03:21,323
For our second goal,

64
00:03:21,347 --> 00:03:25,220
to reduce the use of expensive medical
imaging technologies to screen patients,

65
00:03:25,244 --> 00:03:28,093
we started with a standard,
white light photograph,

66
00:03:28,117 --> 00:03:32,482
acquired either from a DSLR camera
or a mobile phone, for the patient.

67
00:03:32,506 --> 00:03:34,974
Then remember those
billions of information packets?

68
00:03:34,998 --> 00:03:38,498
We overlaid those from
the medical image onto this image,

69
00:03:38,522 --> 00:03:41,474
creating something
what we call a composite image.

70
00:03:41,498 --> 00:03:44,776
Much to our surprise,
we only required 50 --

71
00:03:44,800 --> 00:03:46,133
I repeat, only 50 --

72
00:03:46,157 --> 00:03:50,022
of these composite images to train
our algorithms to high efficiencies.

73
00:03:50,704 --> 00:03:52,029
To summarize our approach,

74
00:03:52,053 --> 00:03:55,220
instead of using 10,000
very expensive medical images,

75
00:03:55,244 --> 00:03:58,252
we can now train the AI algorithms
in an unorthodox way,

76
00:03:58,276 --> 00:04:02,529
using only 50 of these high-resolution,
but standard photographs,

77
00:04:02,553 --> 00:04:05,070
acquired from DSLR cameras
and mobile phones

78
00:04:05,094 --> 00:04:06,617
and provide diagnosis.

79
00:04:06,641 --> 00:04:07,807
More importantly,

80
00:04:07,831 --> 00:04:11,029
our algorithms can accept,
in the future and even right now,

81
00:04:11,053 --> 00:04:13,840
some very simple, white light
photographs from the patient,

82
00:04:13,864 --> 00:04:16,593
instead of expensive medical
imaging technologies.

83
00:04:17,165 --> 00:04:20,228
I believe that we are poised
to enter an era

84
00:04:20,252 --> 00:04:22,167
where artificial intelligence

85
00:04:22,191 --> 00:04:24,720
is going to make an incredible
impact on our future.

86
00:04:24,744 --> 00:04:27,202
And I think that thinking
about traditional AI,

87
00:04:27,226 --> 00:04:29,998
which is data-rich but application-poor,

88
00:04:30,022 --> 00:04:31,569
we should also continue thinking

89
00:04:31,593 --> 00:04:34,601
about unorthodox artificial
intelligence architectures,

90
00:04:34,625 --> 00:04:36,569
which can accept small amounts of data

91
00:04:36,593 --> 00:04:39,539
and solve some of the most important
problems facing us today,

92
00:04:39,563 --> 00:04:40,801
especially in healthcare.

93
00:04:40,825 --> 00:04:42,000
Thank you very much.

94
00:04:42,024 --> 00:04:45,833
(Applause)