Press "Enter" to skip to content

Train and Treat: Cognitive Machines for Medicine | Martin Hedlund & Armando Vieira | TEDxNorrköping


two weeks ago and 31 years back the

Swedish Prime Minister all of pal man

was killed it was shut down in a street

in Stockholm and the biggest police

force ever started to chase the murderer

and look for the gun two months later

the big newspaper two months later the

gun was not found and a big newspaper

came to us and wanted to know if we

could find the gun with our new image

computer so they arranged an experiment

they put a similar gun in a park same

Park as the gunman ran through and then

they took a helicopter photo of that

pork and gave to us and they please find

this gun in this park so we put together

at their time and Al Green and in a

square e in the park we found five

strong responses and in one of those

responses that there was the gun it

showed up there so the headline next day

was computer finds the murder weapon and

it took only 15 seconds but it wasn’t

the murder weapon I can tell you that

at that

we put together the algorithms manually

in a way we created a gun filter that

looked for objects that were banded had

the same shape and size as the gang and

then we searched a an image for that

today we have these artificial

intelligence technology and a special

technique called deep learning and that

is that you build a network and then you

train it with a lot of images and those

images when you train this network it

learns of course and it can turn out if

you recognize persons then in the

beginning the network works not so good

maybe so it detects Hughley instead of

Mary so then an error back propagated in

the network and it updates but that’s

the way it learns so and then it gets

better and better so and it turns up to

a sufficient level and then we are happy

with it then it has built up low-level

features like edges lines mid-level

features like noses eyes muscles and

high-level features like phantom

immediacy and like a dictionary it puts

all this together so it can recognize

persons that it has trained on so then

you can deploy this network and put in

another image that the network has not

seen before and it will predict it

correctly because it has the right

features now we go from an image into a

text and the latest algorri now takes

many objects in in an image and produce

a whole sentence or a description of the

image

and actually all the really latest

research can also put in description and

then go sit synthesize an image and this

is actually communication with images

and I think that is very important in

the future

because we humans communicate very much

with images and we need also these

images in medical area that I will talk

about now we have a lot of images it’s

almost too much data already to handle

and the data it doubles every five years

it’s like a big explosion here so no

humans can handle all this data so we

this is an example of very big data and

that is a pathology image pathologies

technique the image area where you look

at tissues and to find cancer and today

there are arranged competitions for

different research groups and companies

to get smarter algorithms and all of

them get the same data it’s like World

Championships or big tournaments and we

participated in one of those challenges

just a few months ago so it’s really new

things and we succeeded to get a very

good score there a top score and then

our task was to from this pathology

image which is really big I will show

here

TV set high-definition TV set then you

can imagine how much of the image we can

see in this small TV so it’s a huge

amount of data and we get several

hundreds of those data to analyze and

then we find the tumor and then in each

then we sample the image in the tumors

most invasive areas and in high

resolution there we have the secret

because there is cells that are dividing

that’s called mitosis and we look at the

mitotic activities there so we count we

detect the mitosis and those are more

difficult to detect and pistils organs

that we did in 30 years ago and the

number of mitosis that gives us the

aggressiveness of the cancer so if you

look at this image it’s very hard to see

which ones are false or which ones are

true mitosis here for a human so this is

the kind of problems that we are working

with and in the contest we achieved

similar score as a medium pathologist –

and I believe in the future we will beat

pathologist all our machines will be

much better so in combination with

pathologist and machines we can perform

much better so who will make the right

choice in the future will it be doctors

pathologist or will it be machines or

will it be

[Music]

yes research group in United States they

trained a pigeon to detect samples from

breasts and they could be trained to

discriminate between malignant and

benign a cancer benign samples and it

did very quickly and only 25 days it

reached a 90% correct level but there

are problems with having pigeons doing

this because there are regulatory issues

in this and also it’s harder to

communicate with the video so machines

is better I believe and machines can

also help us to interpret complex data

and I give you one example of that here

you can see a pregnant lady scanned with

a three-dimensional ultrasound scanner

and you to the left you can see the

echoes that are reflecting from the

fetus it’s not so easy to interpret in a

good way and get a relation to this but

if you see it in this way it’s the same

data but it has been transformed then

it’s much easier to get a relation to

this right you see it’s a happy foetus

they can be happy before birth also and

I can tell you this baby is now born and

it’s a colleague of mine so and she is

very happy even after birth so cognitive

machines have a great potential they can

handle big and complex data so we can

get

effective healthcare and that will get

better treatment of patients but of

course we need to communicate with

machines in a good way and our Monda

imagination probably is the most

important word that defined humanity it

is through imagination that we discover

science we invent technology and we

create art however imagination is not

the first word that comes to our mind

when you think about machines machines

are machines we don’t talk to a washing

machine to ask the meaning of life just

what the machine to wash our clothes and

you want them to be predictable the same

inputs the same outputs

that’s what machines are for no

unpredictability no noise no imagination

we get surprised if the machine gets

imaginative because we are not expecting

to be imaginative

however machines have evolved a lot in

the recent years now we are machines

they can drive cars machines that can

beat humans in go I have to take to

remind you that this game is very

challenging the number of possibilities

is higher than the number of atoms in

the universe

so there is no way a machine can use

brute force to play this game they use

to rely on something we call intuition

what is intuition intuition is a gut

feeling is somehow we know that the way

is right even if you don’t know exactly

that it is the right option so how does

this machine learn most people who think

that behind these clever machines are

also clever engineers that designed

thousands of lines of code to make this

machine really clever I am one of these

engineers and unfortunately I’m not that

clever

so I have to rely on

techniques one of them was already

mentioned here with the pigeons is

supervised learning in supervised

learning is very easy you just provide

inputs and outputs image and foot or

image and levels there is one other type

of learning is called unsupervised

learning and supervised learning is a

little bit more sophisticated because

you don’t have labels you just have data

is pretty much what babies or young

children do when they start walking they

don’t know the equations of Newton there

were the last days they know how to

learn how to walk there’s one third type

of learning called reinforcement

learning that is pretty much used for

games reinforcement learning is more

tricky because you don’t know don’t have

the model of the world you have to mule

the model of the world and you have to

find a way to get a reward

that can take one day one year or your

lifetime it’s pretty much our lives you

don’t not know what is the you have to

find the path to the reward however I

think there is one for type of learning

that is very powerful

it’s called socialization we humans do

it all the time we are social beings and

how these socialization works it works

for trust Trust is the glue that allow

us to build sophisticated societies that

can not compete with other type of

society from animals without it we will

collapse

however Trust is very unique

characteristics you cannot code it it’s

impossible to cook out trust into a

machine you cannot upload it and you

cannot trade it however Trust is not the

first word to think when we thinking of

machines how can you trust these things

we are after they are smarting us there

are these triangles I don’t believe this

scenario will happen but there’s also

other scenario that most people think

about machines is where they become so

sophisticated that we don’t need to

think they have all the data they have

all the information they are very

powerful and they will make the decision

for us I

think there is one further option that I

think is the most interesting one is

that you get intimate with machines but

intimate I don’t mean we kiss them or

sleep with them I mean we get in natural

with machines because today our

relationship is not natural is very

artificial the moment your car start

talking with you start wondering is it a

car and you go home and your car is in

the street is a kind of loan it is is it

is it getting codes so you unconsciously

put your country six to the humans into

machines but this is not a very natural

relationship so there is one very

important key element in this that I

want you to retain most people thought

about how can you make sure these

machines are behaving properly not going

to be can mean I will have to say that

citing david home that the rules of

morality are not a natural consequence

the reason why other words it means you

cannot code ethics into machine I repeat

you cannot code ethics into machine so

forget about these projects of making

machine safe the machines you become as

safe as we are by interacting with us

the same thing is it your son or your

daughter they will become good or bad

depending of your education so this will

raise many questions like who these

machines have capability in kinds of a

medical scenario decisions of making

life-and-death decisions should they be

legal to have a legal entity should they

be allowed to the patient to take a role

in the decision and what happen is

something goes wrong we put the Machine

into prison

of course our intuition breaks I don’t

have answers for these questions however

should start thinking about them because

in the end we are evolving from

cognitive machines to social machines

you relate to the amputee might try to

relate with humans or animals but the

machines are not humans they are

different so we I’m not completely

convinced that machines will ballerinas

but you’ll be among us so forget about

these

war between men and machines we should

not fear them and we should not ignore

them

we should bind them together with us to

make a better future and to evolve into

a new species to create a better world

thank you very much

[Applause] [Music]

Please follow and like us: