1
00:00:01,280 --> 00:00:05,136
Computer algorithms today
are performing incredible tasks

2
00:00:05,160 --> 00:00:09,896
with high accuracies, at a massive scale,
using human-like intelligence.

3
00:00:09,920 --> 00:00:13,856
And this intelligence of computers
is often referred to as AI

4
00:00:13,880 --> 00:00:15,736
or artificial intelligence.

5
00:00:15,760 --> 00:00:19,960
AI is poised to make an incredible impact
on our lives in the future.

6
00:00:20,880 --> 00:00:24,816
Today, however,
we still face massive challenges

7
00:00:24,840 --> 00:00:28,336
in detecting and diagnosing
several life-threatening illnesses,

8
00:00:28,360 --> 00:00:30,720
such as infectious diseases and cancer.

9
00:00:32,000 --> 00:00:34,296
Thousands of patients every year

10
00:00:34,320 --> 00:00:37,120
lose their lives
due to liver and oral cancer.

11
00:00:37,880 --> 00:00:40,576
Our best way to help these patients

12
00:00:40,600 --> 00:00:44,920
is to perform early detection
and diagnoses of these diseases.

13
00:00:45,880 --> 00:00:50,040
So how do we detect these diseases today,
and can artificial intelligence help?

14
00:00:51,920 --> 00:00:55,576
In patients who, unfortunately,
are suspected of these diseases,

15
00:00:55,600 --> 00:00:58,256
an expert physician first orders

16
00:00:58,280 --> 00:01:00,896
very expensive
medical imaging technologies

17
00:01:00,920 --> 00:01:05,016
such as fluorescent imaging,
CTs, MRIs, to be performed.

18
00:01:05,040 --> 00:01:07,336
Once those images are collected,

19
00:01:07,360 --> 00:01:11,880
another expert physician then diagnoses
those images and talks to the patient.

20
00:01:12,520 --> 00:01:15,976
As you can see, this is
a very resource-intensive process,

21
00:01:16,000 --> 00:01:20,416
requiring both expert physicians,
expensive medical imaging technologies,

22
00:01:20,440 --> 00:01:23,536
and is not considered practical
for the developing world.

23
00:01:23,560 --> 00:01:26,920
And in fact, in many
industrialized nations, as well.

24
00:01:27,760 --> 00:01:30,640
So, can we solve this problem
using artificial intelligence?

25
00:01:31,840 --> 00:01:35,896
Today, if I were to use traditional
artificial intelligence architectures

26
00:01:35,920 --> 00:01:37,136
to solve this problem,

27
00:01:37,160 --> 00:01:38,616
I would require 10,000 --

28
00:01:38,640 --> 00:01:42,656
I repeat, on an order of 10,000
of these very expensive medical images

29
00:01:42,680 --> 00:01:44,056
first to be generated.

30
00:01:44,080 --> 00:01:46,976
After that, I would then go
to an expert physician,

31
00:01:47,000 --> 00:01:49,496
who would then analyze
those images for me.

32
00:01:49,520 --> 00:01:51,616
And using those two pieces of information,

33
00:01:51,640 --> 00:01:55,296
I can train a standard deep neural network
or a deep learning network

34
00:01:55,320 --> 00:01:57,456
to provide patient's diagnosis.

35
00:01:57,480 --> 00:01:59,216
Similar to the first approach,

36
00:01:59,240 --> 00:02:01,383
traditional artificial
intelligence approaches

37
00:02:01,407 --> 00:02:02,856
suffer from the same problem.

38
00:02:02,880 --> 00:02:07,440
Large amounts of data, expert physicians
and expert medical imaging technologies.

39
00:02:08,320 --> 00:02:12,616
So, can we invent more scalable, effective

40
00:02:12,640 --> 00:02:15,936
and more valuable artificial
intelligence architectures

41
00:02:15,960 --> 00:02:19,016
to solve these very important
problems facing us today?

42
00:02:19,040 --> 00:02:22,336
And this is exactly
what my group at MIT Media Lab does.

43
00:02:22,360 --> 00:02:26,216
We have invented a variety
of unorthodox AI architectures

44
00:02:26,240 --> 00:02:29,416
to solve some of the most important
challenges facing us today

45
00:02:29,440 --> 00:02:31,640
in medical imaging and clinical trials.

46
00:02:32,480 --> 00:02:35,536
In the example I shared
with you today, we had two goals.

47
00:02:35,560 --> 00:02:38,536
Our first goal was to reduce
the number of images

48
00:02:38,560 --> 00:02:41,816
required to train
artificial intelligence algorithms.

49
00:02:41,840 --> 00:02:43,936
Our second goal -- we were more ambitious,

50
00:02:43,960 --> 00:02:47,696
we wanted to reduce the use
of expensive medical imaging technologies

51
00:02:47,720 --> 00:02:48,936
to screen patients.

52
00:02:48,960 --> 00:02:50,160
So how did we do it?

53
00:02:50,920 --> 00:02:52,136
For our first goal,

54
00:02:52,160 --> 00:02:54,216
instead of starting
with tens and thousands

55
00:02:54,240 --> 00:02:57,256
of these very expensive medical images,
like traditional AI,

56
00:02:57,280 --> 00:02:59,336
we started with a single medical image.

57
00:02:59,360 --> 00:03:03,136
From this image, my team and I
figured out a very clever way

58
00:03:03,160 --> 00:03:05,896
to extract billions
of information packets.

59
00:03:05,920 --> 00:03:09,616
These information packets
included colors, pixels, geometry

60
00:03:09,640 --> 00:03:12,176
and rendering of the disease
on the medical image.

61
00:03:12,200 --> 00:03:16,536
In a sense, we converted one image
into billions of training data points,

62
00:03:16,560 --> 00:03:20,096
massively reducing the amount of data
needed for training.

63
00:03:20,120 --> 00:03:21,336
For our second goal,

64
00:03:21,360 --> 00:03:25,216
to reduce the use of expensive medical
imaging technologies to screen patients,

65
00:03:25,240 --> 00:03:28,096
we started with a standard,
white light photograph,

66
00:03:28,120 --> 00:03:32,456
acquired either from a DSLR camera
or a mobile phone, for the patient.

67
00:03:32,480 --> 00:03:34,936
Then remember those
billions of information packets?

68
00:03:34,960 --> 00:03:38,496
We overlaid those from
the medical image onto this image,

69
00:03:38,520 --> 00:03:41,040
creating something
that we call a composite image.

70
00:03:41,480 --> 00:03:44,776
Much to our surprise,
we only required 50 --

71
00:03:44,800 --> 00:03:46,136
I repeat, only 50 --

72
00:03:46,160 --> 00:03:50,000
of these composite images to train
our algorithms to high efficiencies.

73
00:03:50,680 --> 00:03:52,016
To summarize our approach,

74
00:03:52,040 --> 00:03:55,216
instead of using 10,000
very expensive medical images,

75
00:03:55,240 --> 00:03:58,256
we can now train the AI algorithms
in an unorthodox way,

76
00:03:58,280 --> 00:04:02,536
using only 50 of these high-resolution,
but standard photographs,

77
00:04:02,560 --> 00:04:05,056
acquired from DSLR cameras
and mobile phones,

78
00:04:05,080 --> 00:04:06,616
and provide diagnosis.

79
00:04:06,640 --> 00:04:07,856
More importantly,

80
00:04:07,880 --> 00:04:11,016
our algorithms can accept,
in the future and even right now,

81
00:04:11,040 --> 00:04:13,856
some very simple, white light
photographs from the patient,

82
00:04:13,880 --> 00:04:16,320
instead of expensive
medical imaging technologies.

83
00:04:17,120 --> 00:04:20,216
I believe that we are poised
to enter an era

84
00:04:20,240 --> 00:04:22,176
where artificial intelligence

85
00:04:22,200 --> 00:04:24,736
is going to make an incredible
impact on our future.

86
00:04:24,760 --> 00:04:27,216
And I think that thinking
about traditional AI,

87
00:04:27,240 --> 00:04:30,016
which is data-rich but application-poor,

88
00:04:30,040 --> 00:04:31,576
we should also continue thinking

89
00:04:31,600 --> 00:04:34,616
about unorthodox artificial
intelligence architectures,

90
00:04:34,640 --> 00:04:36,576
which can accept small amounts of data

91
00:04:36,600 --> 00:04:39,536
and solve some of the most important
problems facing us today,

92
00:04:39,560 --> 00:04:40,816
especially in health care.

93
00:04:40,840 --> 00:04:42,056
Thank you very much.

94
00:04:42,080 --> 00:04:45,920
(Applause)