Player is loading...

Embed

Embed code

Transcriptions

Note: this content has been automatically generated.
00:00:00
uh with former actually
00:00:03
thank you very much for uh for inviting me the last uh it's always a pleasure to go uh
00:00:08
speak in front of an internal does not been doing a lot leave his thanks to you
00:00:13
so uh let's see what uh oh what um i can tell you
00:00:17
about the perspective of the radiologist looking at a high and
00:00:23
just to give you um uh the atmosphere in which we have been living the for
00:00:29
the for a past for four years or so uh let's look at this uh
00:00:35
video
00:00:38
so
00:00:42
it's it's it's it's it's
00:00:55
ah she is just it's it's
00:01:06
i i ah
00:01:15
ah
00:01:16
so a disguise actually uh uh one of the ah constant beacons in that computer
00:01:22
science it's a a works uh it cool also teaches at university of toronto
00:01:27
and and what is says we should stop of a training rejoice
00:01:31
we don't need anymore george's because yeah we'll take over
00:01:34
is uh is not the only one to think so uh now if you look at the business side
00:01:38
of things uh people like famous influence there's like a few not cost lab a venture capitalist
00:01:44
uh was a given an interview in this paper and i start reading it for me was like roller coaster so uh
00:01:50
so first line he says that the role of the radio this would be obsolete in five years and i wasn't too doesn't seventeen i think
00:01:57
um so i started the to count the the years and i almost pick up my
00:02:01
phone to call thomas please take me as a reason to remotes elegy i
00:02:05
and then i kept reading and the next the sentence was a machines would replace eighty
00:02:10
percent of doctors and i was like okay so what do we do now
00:02:14
and uh actually the journal this went to san says stanford
00:02:18
and san francisco and ask a few radiologist or what
00:02:21
they think of all of that a. i. n. and so on and the radio to say that their jobs
00:02:27
we only become more important with air out so i start to breathe again this might come back on my face
00:02:33
but nevertheless this is what we've been hearing over and over
00:02:36
again a i will threaten other jobs off radiologist
00:02:41
so what i'm gonna be telling you in the lesson next uh fifteen twenty minutes
00:02:47
is we're gonna try to see if the reason room for optimism or not a former with all this perspective
00:02:53
in order to understand this or to answer this question will see what do what a i can do
00:02:58
for us as we all just how we would be impacting our work flow our daily work full
00:03:04
and it's called without saying that there are limitations and challenges ahead
00:03:08
of us as well in integrating i now work flow
00:03:11
and i will go a touch upon that were briefly as well let's start a little bit of um historical perspective
00:03:19
a religious have been fully your steak is actually trying to find a ways
00:03:24
for computers to help them improve the way they work be more efficient
00:03:29
being a a more accurate this is nineteen eighty nine uh a peep a paper
00:03:35
on the problems in applying expert system technology to the interpretation of requests
00:03:41
and the problems they had at the time were the very different from from what we have now
00:03:46
there were asking themselves how the ritual utilities to fit should phrase
00:03:51
the the abnormality that b. c. to putting in the computer in the right way
00:03:55
for the computer to tell them the right diagnosis so that was very primitive
00:03:59
thirty years ago let's jump back in time ten years ahead of
00:04:03
that so nineteen ninety seven that was um the paper
00:04:07
written in one of the most influential journals uh in our field uh
00:04:11
uh that was called can machine learning or can machines to learn
00:04:15
to interpret read about so that was already uh the question that was asking that they're asking them selling a twenty years ago
00:04:22
and in this paper the uh our asking does this mean they're they're risking expect to be replaced
00:04:27
but network machines in a new feature answer absolutely not to that was already optimistic by then
00:04:33
but they were saying that uh in the next decade we will
00:04:36
likely produce network based system that eight we deal or just
00:04:39
by previewing the images in and helping them basically uh improve the work flow
00:04:45
that was nineteen ninety seven now let's move on to in two thousand
00:04:49
seventeen to present times uh what is what is happening today
00:04:54
the if you look at the literature theory was this a review
00:04:58
paper uh that looked at three hundred over three hundred papers
00:05:02
in part meant um where do you look for a publications on a on a machine
00:05:08
learning deep learning in medical imaging analysis and this is uh what they found basically
00:05:16
an exponential increase in the number of papers in that field and exponential means basically
00:05:21
the that was options as to doesn't seventeen most of the papers were
00:05:24
published two dozen fifteen sixteen so the next free day you're sorry
00:05:29
every modalities are concerned as you can see him rise city and so on and
00:05:33
so forth different uh techniques segmentation detection
00:05:37
classifiers uh i've been a used
00:05:40
and a surprisingly pathology was first in the top of a of
00:05:44
the fields were most affected pathology also have images microscopes
00:05:49
but a different field in an an radiology brain in someone absurd for me
00:05:54
i was surprised to see that blown imaging was not so highly
00:05:57
ranked and actually i did uh i just finished a right to review one on m. s. k. a. i. m. es get radiology
00:06:02
and if you look at the papers that were published up to the end of two thousand eighteen there was only thirty papers
00:06:08
i know i miss kate radiology so up arthritis detection and so on so forth
00:06:13
um so it it it's it's lighting a little bit behind a i. e. maybe we'll have time to discuss the reason why
00:06:21
um so it has been a hard topic it still is a all the more our
00:06:26
topic today since the advent of deep learning in and in the past few years
00:06:30
and if you wanna look at how the radiologist will is seeing the impact of a i nice work full
00:06:37
let's look at what we do in our daily lives so we receive a request from a clinician
00:06:42
we uh manage this request we look at it decide what to do with them a lot
00:06:46
is correct or not we schedule a we schedule the appointment and then reproduce the images
00:06:52
um there's a few steps here to uh puerto calling deciding how to do how to perform the images and so
00:06:58
on and so forth we interpret the image that's produced make a report transmit to see a couple franklin issue
00:07:07
first step managing to request an schedule like
00:07:10
does the patient has allergies to a conference media that's something
00:07:14
that's some clinicians uh forgot to tell us about
00:07:18
so we have to be certain above of that before we start the the main
00:07:21
uh the examination before we choose maybe switch modalities be depending on the answer
00:07:26
patient as a pacemaker and so on and so forth piece can become completely automate size especially using
00:07:33
a language a language classifiers so that's something we
00:07:37
are actually hoping for and and we're waiting
00:07:40
um then we need to decide if i have to do a more i have the of the ne what's sequences what
00:07:46
planes will uh if i need it contrasts midi or not
00:07:50
instruments of course i first years of question last myself
00:07:53
and and the answers is based on what uh i'm looking for so what
00:07:57
i'm looking for a and unfortunately the conditions are very good at hiding
00:08:01
us the not the correct information that important information for us at another work they want to test us every time the ice 'cause it
00:08:08
an exam so um so this is also for us a very
00:08:12
important uh in a i can have a huge impact
00:08:15
in terms of uh decided to helping us decide how to protocol in can be automatic so if you uh
00:08:22
a look at this algorithm uh it did look at the full text of a clinical indications of the study
00:08:28
and they decided whether now uh injection ivy contrast was needed or not in
00:08:33
eighty percent of the cases why would the accuracy of eighty percent
00:08:37
and uh other uh applications uh exist uh that's
00:08:42
something we are actually hoping for a um
00:08:46
i have acquired the digits they're men now i need to reconstruct the
00:08:51
raw data so i put my patient in in the m. r.
00:08:54
and i have raw data to uh to play with and to to to build an image from
00:09:00
if you have a meaning if you're aiming at having such an image this is or
00:09:05
crawl view of an e. m. r. i. uh this is a zoomed in view
00:09:10
you can now using generally if a i models uh acquired
00:09:17
this which is a much faster law requisition a image
00:09:21
and then from band from that build this so using models a.
00:09:27
i. models you can be very close to a high resolution
00:09:30
uh image so that's quite something for us it's a it's a it's something that's very very interesting in terms
00:09:37
of uh increasing the resolution over images becoming faster right acquiring them and so on and so forth
00:09:43
it was a very interesting papers published um they looked at how how does generative models can
00:09:49
can switch basically ten can get the information from one image and and predict the
00:09:55
another sequence for example let's start with the non fat suppressed to to sequence of the spine here
00:10:01
you see that the fact is not suppressed a that would be the fast
00:10:05
suppressed image that you acquire this is the output of the algorithm so
00:10:10
basically from that it constructs that it predicts that which is very very similar
00:10:14
to our target it's not perfect but it's extremely similar even crazier
00:10:21
you give them a really rough
00:10:25
people in more
00:10:27
i could not believe it i was uh i was shocked uh it a date from that they did this
00:10:34
so um i don't know i'm i'm really curious to see what else they can do a it's it's
00:10:39
it's not perfect if you look at this c. area here it's not exactly the same but it's it's
00:10:43
it's the the potential is enormous enormous and and actually the question is we we actually need
00:10:50
to look at these images uh or can we just uh um deal with the raw
00:10:55
data but that's that's for the very like long term for sure i guess
00:11:00
image interpretation that's the next step of my of my daily work flow uh i have
00:11:04
the image i need to uh decide what uh what the diagnosis is and and
00:11:10
lesion detection is the first thing that comes to mind when people think of a i apply to radiology
00:11:16
and that machines can do better than than human beings and for most
00:11:20
of it like the asian detections typical example would be fracture detection
00:11:24
uh the there is an error a misinterpretation of average emergency missing functions by
00:11:30
forty seven percent so they feed the algorithm with the experts analysis
00:11:36
the all rhythms apply to relax the uh and they perform the algorithms perform
00:11:41
much better than the average commission so that's something very very promising for
00:11:45
to implement uh to increase the the the efficiency in the performance of of all the
00:11:50
doctors who are not experts in those fields who have to deal with these images
00:11:55
uh other applications you for example you have a stack of mammograms you're doing a
00:11:59
mammogram uh uh a console today and we have forty mammograms to look at
00:12:04
most of the mammograms are actually normal in a in a hopeful i mean thankfully so
00:12:10
um but if there's one had normal mammogram you prefer or
00:12:15
looking at it while you're not digesting relaunch right so
00:12:20
you can pretty digest yeah i can pretty digest those images put it at the top of the list
00:12:26
and you know that this one probably is abnormal so i'll go look at it when i i
00:12:29
just have coffee and and i'm awake basically to look at it and not miss it
00:12:33
that's something that's very important um you have a stack of c. t.s acquired at the emergency room
00:12:39
and uh one of them has a a a has an abnormality
00:12:44
and and before uh if if you wait two hours uh it's it's
00:12:48
it's it's it's a shame so you can have this image flight
00:12:51
put at the top of your list and take care of the patient very urgently so these are the applications that uh we um
00:12:58
we see but image interpretation does not only mean looking and image and making a
00:13:02
diagnosis it uh uh this is a patient i i had looked at
00:13:07
two days ago fifty five year old woman with long burping this is the information i had
00:13:11
and i look at this i saw this no idea what it out with it
00:13:15
so because i know that clinicians they want to hide information from us
00:13:19
i went to look at digital records and i find out information
00:13:22
of interest information breast carcinoma dispersion as history of arsenal
00:13:27
i forgot to tell us um and actually the pain was will be lower than when nice what
00:13:32
that what what is this but i know also that it's not always reliable but the
00:13:36
commission tells us the l. four or five in the hands of remote all this isn't the
00:13:39
same as a in the hands of a of i don't know a dermatologist so uh
00:13:46
more questions and uh and i see i think okay maybe i need to look
00:13:50
at uh at the previous uh data and see if this was already present
00:13:54
i look at the list of of the images thirty exams for this patient in my pack system
00:13:59
i assess part one that could be interesting for me to put city from today two years ago
00:14:04
i look at it and go down and this something here that might be what i'm
00:14:09
looking for but i can't really compare those two images i need to have
00:14:12
said it'll images to do that i need to go out was that close exam
00:14:16
real open it again to re formatted in uh in this as a plane
00:14:19
reformatting scrolling panning zooming windowing five minutes later i have this image and
00:14:24
there is something you are not so not sure what it is
00:14:27
but now i can have the information at hand to decide what to do to another examination or follow up or whatever
00:14:33
it took me twenty five minutes to think about this case and all of that
00:14:37
could have been ultimate ties with yeah i and that this is something uh
00:14:41
uh i'm not even talking about um a quantitative him a analysis of images this
00:14:46
is something we need to do with you wanna do proper research or even
00:14:50
that's with our patients probably we need to add quantitative what
00:14:54
the same a quantitative a day that you know reports
00:14:57
time consuming cumbersome nobody does it uh the t. two mapping to look at
00:15:03
the composition of college is all this that i had to go through
00:15:06
to build this image you and get a value for uh to to mapping fifteen minutes nobody that
00:15:12
one image every two seconds every working day for an entire this is uh the live coverage
00:15:17
or just and i'm sure they must be a better way to do this uh a
00:15:23
soul that's why probably if you look at the uh public opinion in these days if you
00:15:27
look at the if you wanna know the public opinion something you look at treats
00:15:31
you look at two years three or six hundred and five squeeze a by four
00:15:35
and seven users over one year uh that included a i in imaging
00:15:41
um most of them are very positively we already recording uh considering
00:15:48
yeah i i never george's so i eighty eighty percent
00:15:51
and not very few we say that a i would replace with all just
00:15:56
with vision seemed work for improvement are the main uh uh things that come up and limitations of
00:16:02
challenges and i want to briefly touch upon that uh because i think it's important so
00:16:08
two risks with a i we already talked about it earlier just to today is really depending on on how you
00:16:14
feed the algorithm if you don't have good or if you have bias input you gonna end up with disaster
00:16:22
there is it's very easy down days to build an algorithm that can
00:16:24
label uh images you'd feed them with images that are labelled
00:16:28
you teach them how to label the images and you get an answer for new set of data out new sort of images
00:16:35
they did this test with a a standard uh yeah i which was exposed to images from google
00:16:42
and then uh they build the psychopath i'll go with them
00:16:47
which was fed by images taken from the dark what okay
00:16:53
and then they they asked them to a label this image for example
00:16:58
the standard a. i. c. is a group of birds sitting on top of up to re backed branch why not
00:17:10
look at this one
00:17:16
so depending on the data you feed the algorithm the result can be
00:17:21
extremely different right not to say that it's this is a paper or thirty first to generate it doesn't nineteen yesterday
00:17:28
if for example for patients to worse after organ
00:17:31
transportation or after receiving chemo friends this cancer
00:17:36
machine learning algorithms may conclude such patients are less likely to benefit from for the treatment
00:17:41
and we commend against it so again if you take it data and you can buy
00:17:46
sits in some sort of way you end up with a disaster um liability issue
00:17:55
somebody has to take responsibility for for with the machine says and as of today i don't know
00:18:01
many people in this room who wants to go on the flights with no human pilots
00:18:06
uh in the flat i know we know that's last last uh aeroplanes can fly at all
00:18:11
uh we actually probably need somebody sitting in this the seat and and uh
00:18:16
and that's very important to to have to convert the patients and the
00:18:20
radiologist the pilot can talk to the to the to the
00:18:23
'kay can be the the gap between the bridge the gap between knowledge
00:18:27
and and uh and and reality you can if there's a turbulence
00:18:31
i wanna hear if human voice telling me there's a turbulence i wanna know all that is located the plane is doing this
00:18:37
and and and i wanna hero human voice a reassuring human voices
00:18:41
you and into the pad doesn't give me this information i
00:18:45
stress for like fifteen minutes like what's going on so it's something that really need the need to uh
00:18:50
when we have to have that in mind i'm gonna go over that very quickly because it's so
00:18:56
in terms of health care artificial intelligence we we all
00:19:01
this we look at as augmented intelligence we needed
00:19:05
to improve our fifties that's something very uh very uh important and that's end up on this not
00:19:13
thank you very much higher
00:19:17
hi thank you very much soul scary in exciting the same time
00:19:25
so when we see a whole the possibility good summaries
00:19:28
so it's like a used to hide in the yellow
00:19:30
with drinking decided whatever sell ah questions please
00:19:38
when i got to move on until i get half he's one question from the floor
00:19:43
yeah but it that's cheating
00:19:49
i just um put some protects you know my tastes in in m.
00:19:53
arrives any how far we this be nice to have huh
00:19:58
automatically it can be done to morrow but like that but again or is it not of interest could be done if you have the right data
00:20:05
and you have the proper labelling okay depends on who you check to label data
00:20:10
if you wanna pay an expert like me very expensive yeah or if you
00:20:13
wanna have somebody your kid for example do that uh after school
00:20:17
so it's really is a lot of issues in terms of like yeah during the right data we already talked
00:20:21
about this morning but when is m. s. k. compared to um um colour g. in automatic image recognition
00:20:28
why or where where are we sell like i imagine the demo applications now for cancer
00:20:34
i'm not exactly so one of the main hourly and yeah yeah that that
00:20:39
again so like one of the next things that can uh it
00:20:41
being implemented is the detect automatic detection of long nodules
00:20:45
for a screening a lung cancer and this is something like probably it's gonna happen in within five years or so
00:20:52
but for that you need registries approaches repositories of of patients with follow up and and we know
00:20:58
that this patient have this not will develop into cancer so we have to which respective analysis
00:21:03
any any cancer lung cancer with this it this is something that we have to do yeah
00:21:07
already and and it's not doesn't takes all so much effort to to an ally set
00:21:12
these are the areas where we will be forced to using a i
00:21:17
i i i whenever you go into deeper or a complexity
00:21:21
it's gonna in my opinion we not there in the for ten twenty years
00:21:25
but again a eyes not just looking at images in diagnosing a eyes also a
00:21:30
language recognition language analysis and this can be a lot of help for us
00:21:36
so what is the taken busy age i pro all the calls
00:21:41
of knowledge i'm this life sure you don't agree i mean
00:21:45
i wasn't clear enough or you try to push a few to be too far away when you hear twenty years
00:21:51
twenty is was so we think it happens but she was gone to a computer is some other things
00:21:56
goes so fast it's the years you will flying plane without any point that's
00:22:01
a psychologically speaking almost okay but they would have a nice voice telling
00:22:04
you everything is fine yeah probably i know it's a computer even you cheer
00:22:08
we just i'm not stupid i know it it's okay easy but
00:22:12
for the which at you to smoke something in the plane so the okay
00:22:16
to answer the question i think the question is not if i'm pro
00:22:19
or con uh there was a paper in two thousand and this month
00:22:23
in radiology which is our our journal saying resisting is for china
00:22:28
it is no i question it's gonna happen so i i it's not offer made to
00:22:32
resist something that is gonna happen so the question is how we wanna make it
00:22:38
uh how wouldn't we wanna use it for our best interest for us for
00:22:43
the medical community in general and mostly formations of course of course
00:22:48
alright thank you very much thank you idea oh one question one okay yes i get one
00:22:57
and thank you patrick for your very interesting a presentation um maybe function is for for
00:23:03
what you you you say yeah we need some collaborative uh work uh to
00:23:08
help uh to support 'em doctors and you in this evolution do you know the
00:23:15
i. e. they any uh people of
00:23:19
a specialist of informatics data management
00:23:22
um i i a. e. engineer who work with doctors like that or what we have that
00:23:28
issue have that or yeah yeah we just hire the person was the king himself uh
00:23:32
uh is actually siemens employee but yes seventy percent of the salary come from the really off department
00:23:38
any as well i mean he also teaches at a pay fail uh has a
00:23:41
whole team around him and his integrate in the centre of biomedical imaging
00:23:46
uh which is about thirty physicist uh developing sequences are and all that so these
00:23:51
people this a whole team of people in the radiology department at issues
00:23:56
and i'm only talking about issues uh working uh you know that
00:24:00
the scientists that we have uh we just bought a
00:24:03
servers uh from the shoes to do that so it's yeah we have we on it right
00:24:09
short questions so in the clinical parks is one of the most important thing
00:24:13
is to compare serial images show you the the food to your room
00:24:17
twenty five minutes to try to find a needle in the haystack show
00:24:21
why is it so with a difficult to just the question i'm asking myself every day
00:24:26
you don't understand why is this is not automatic yeah two dozen nineteen i can talk to my
00:24:32
phone in and make as a you know a task for restaurant tonight any bookings for me
00:24:37
and this a match is not aligned averages are probably it's it's just uh you know it's
00:24:43
one of the issues also like this the business is slow to grow uh uh the the the
00:24:47
business partners they are the legislation around all this whenever you have to look at the image
00:24:54
this so much regulatory stuff to go through before you implement software new software
00:24:59
takes forever takes five years so so in in one of the things that we need
00:25:03
as with george's i was talking to one of the colleagues uh at the break
00:25:07
i don't want to stand up for my computer go to another computer do a software
00:25:12
analysis that that that's not available in my packs go back to my packs and
00:25:15
and reporting with a computer i want everything to be integrated because you need that
00:25:19
integration uh it takes time wriggle two issues and so on and so forth
00:25:25
um i think the diagnostic accuracy of any taste
00:25:31
not only depends on a tuesday tilted discrete
00:25:34
you were diagnosed a gal rules more interpretation aged
00:25:37
eulogy but it depends also on the pretext
00:25:41
um frequent problem and probability in other words the person it's re
00:25:46
for you the patient how to well groomed account for that
00:25:51
well they will have to confirm that uh that's for sure
00:25:57
that's that's a whole you know why would we talked about the us
00:26:00
when you arthritis issue and and uh the uh classification uh
00:26:04
criteria were talking about it at dinner last night at the talk i was giving and is your last year
00:26:09
uh uh it's it's all biased like you know do you use the
00:26:13
criteria that's a lot to classify patients in a in a
00:26:15
specialised environment and you apply those criteria to make a diagnosis in
00:26:20
a normal way uh uh you know normal uh context or
00:26:24
that's why i i personally think the human factors to bear important
00:26:27
to leeds dialogue and before an algorithm can probably not
00:26:33
always replace 'em every year every radiologist i'm convinced
00:26:41
about that yeah i i hope i'm right
00:26:44
thanks i hope for you to so thank you very much time to move on
00:26:55
is the next speaker posters alton could tell it from a dozen university
00:26:59
an interesting because for me to list the conflicting words in the same
00:27:03
titles so that be very good to learn is big that i'll
00:27:08
and rainy bongo rate disease so is like a rate data always big that also had the

Share this talk: 


Conference program

Demystification the digital health world
Thomas Hügle
1 Feb. 2019 · 9:41 a.m.
200 views
Digital lessons learned from musculoskeletal radiology
Patrick Omoumi
1 Feb. 2019 · 11:37 a.m.