Player is loading...
Q&A
Dr Amine Korchi
Monday, Feb. 18, 2019 · 11:43 a.m. · 13m 25s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:03
oh
00:00:09
oh
00:00:13
oh
00:00:17
i uh if you had one your print out one hundred after her
00:00:28
yeah
00:00:34
oh that's an excellent question so definitely if you have one hundred year of training
00:00:40
you have seen many cases you have followed up many cases so we have seen images of patients that
00:00:46
undergoes surgery and then analysis of the tissue then you know all does this image is related to
00:00:52
just kind of cancer is related to this kind of outcome with this treatment so yes and that's
00:00:57
why we g. pruning is trained on large data sets it sitting the human ability to learn
00:01:05
we are not able to learn on ten hundred million images in one year
00:01:15
oh
00:01:19
uh
00:01:25
oh
00:01:32
yeah my impression is it's dependent of the that that's what
00:01:36
we call the experience in professor who is sixty five
00:01:39
your role have seen hundreds if not thousands of not that i know millions of cases he has better experience
00:01:51
oh
00:01:54
e.
00:02:10
so from my perspective my personal perspective i embrace
00:02:13
this movement since years i'm annotating i'm working
00:02:16
on the strategy level with the last cooperate startups from i would say uh the
00:02:22
majority of the radiologist there where kind of if you hear a few years ago and
00:02:27
uh since i would say one two years the radiologist and the radiology department
00:02:32
has begun corroborating because they are slowly accepting the idea
00:02:38
that i i would not replace them at all
00:02:41
but they will augment them we're shifting from here to a collaboration since two years
00:02:50
yeah
00:02:55
i see
00:03:01
i
00:03:04
a a a one
00:03:17
i i i
00:03:25
uh_huh
00:03:28
a a a a a a a
00:03:43
oh yes the from the the answer is straightforward today that
00:03:49
the answer is no the generalisation over different vendors
00:03:54
even different plans of acquisition when we make images of
00:03:58
the brain usually we have to reference planes
00:04:01
what we call c. a. c. p. or in english uh the op was
00:04:04
it so it's approximately we this the slides are oriented like these
00:04:10
i have seen on many application that if you just do a little bit
00:04:16
the algorithm does not segment the brain re why he does not recognise the structure
00:04:21
so this is when the same machine when we go over all their machines
00:04:26
the generalisation face today and this is also
00:04:30
a problem uh that is widespread in health care what we call inter operability
00:04:36
there there are no inter operability each vendor has its own structure outside
00:04:43
it's silent so there is a lot of work to overcome this kind of uh of problem
00:04:50
obviously if you if you do i'll go over it with images
00:04:53
from all vendors at some point it could be generous
00:04:58
but the algorithm does not have a i kissed what i have seen it does not have the knowledge of the anatomy
00:05:05
yeah and this is very crucial he does not have the knowledge of the treaty structure he's processing
00:05:13
usually to the images to the slices and is trying to recognise patterns
00:05:18
that's i believe one of the limitation as a medical doctor we first
00:05:22
train as doctors knowing the human body we know what the anatomy
00:05:26
besides medical imaging designs anything we know how brain is is is from the anatomical structure
00:05:33
that would be one research uh i would say i we are when i was discussing with some research chain to
00:05:40
try to be inspired by this kind of learning learning for
00:05:44
the anatomy maybe the actress actresses of the brain
00:05:47
and then building algorithm that are more aware of this but it
00:05:50
from what i heard today it's quite hard because we're
00:05:54
still in to the pattern recognition so the the the answer is no generalisation is not something given to that
00:06:09
yeah i
00:06:15
yeah
00:06:20
i
00:06:26
uh_huh
00:06:32
uh
00:06:35
i i i a
00:06:44
i
00:06:53
i
00:06:56
i'm not aware of uh this kind of inclusive holistic studies but some
00:07:02
research a unit i can for example talk about i. b. m.
00:07:06
uh the research team a devoted to medical imaging i had a talk with that one of their team a blaster
00:07:13
they have this sense of mean making the natural work flow in the organisation and in the mind of
00:07:19
the radiologist but i'm not aware of a publication maybe exist but i'm not aware of publication looking at these
00:07:26
no for the beneficial effect of replacing one or two
00:07:29
steps on the health care system outcomes and costs
00:07:35
and it would be very hard to obtain this kind of studies
00:07:42
yeah
00:07:45
oh yeah at the last good uh_huh
00:07:58
yeah
00:08:02
that's the whole point uh having go a wide view of the system
00:08:07
from starting point two and point and trying to envision and a i uh
00:08:16
had begun shift how we can transform the system how we can make it more streamlined
00:08:21
at the whole the problem is health care system are very different from one
00:08:26
country to the others and within one country it is highly fragmented
00:08:31
takes with that and uh we have approximately six to eight
00:08:36
or nine high level hospitals for eight million people
00:08:40
in the bay area ah around silicon valley have too many hospitals for the
00:08:44
same amount of people so it gives you a sense of fragmentation
00:08:49
inside and outside healthcare systems
00:08:54
any other questions
00:09:02
oh oh oh oh oh
00:09:16
i work for a while
00:09:22
i i i
00:09:31
yeah
00:09:33
definitely so definitely we come back to the fragmentation of healthcare inter operability proprietary says
00:09:41
then silo data uh but uh it is the same challenge in computer vision
00:09:49
it we have also the same challenge in computer vision we have different
00:09:52
type of images where different type of machine so those challenges
00:09:57
are present in all the healthcare use cases what is and
00:10:01
what i would say this year um is for example
00:10:06
if you tried to souls the automatic scheduling problem
00:10:10
you are less dependent of uh of outside and that'll make knowledge
00:10:15
and you are more dependent on numbers schedules
00:10:21
roll data that you can more easily process
00:10:33
oh i
00:10:40
she
00:10:45
i wish
00:10:53
i i think the best
00:10:56
position people are the clinical that assigns and the uh
00:11:01
d. global example or it's the n. g. h. and b. w. h.
00:11:06
o. clinical that assigns so we'd sit at that assigns institute
00:11:10
which is backed by the hospital they have full access to the to the health care data they have full
00:11:17
access to the expertise and they have in house doctors
00:11:21
in house doctor saying just in house product developers
00:11:25
so they interact daily with the clinicians with the work flow that developing their own prague
00:11:31
proof of concept then they are integrating does proof of concept into a sandbox
00:11:37
very localised clinical teen taking feedback and iterating
00:11:42
dues for me are the best position to bring innovation because they
00:11:47
are they have the height of expertise of that assigns as
00:11:50
jointly with the high level expertise of clinicians in a clinical environment
00:11:56
and this is the key to be close to the clinical environment and then we will see we cannot predict
00:12:02
how fast it will arrive what would be the impact but my impression is uh
00:12:07
the administrative work flow would benefit from a i earlier then image analysis
00:12:20
i i i
00:12:25
a bear has a open to that uh science institute i was
00:12:29
invited to be that a similar talk of two months ago
00:12:33
they are building at that age with anonymous the top and offered the opportunity at least it
00:12:39
is the project for industry startups research team to log into their that i like
00:12:45
and develop their own ago if i'm aware of of burned you know with the engine speed that
00:12:53
but globally uh outing dog but neither is the uh the
00:12:57
u. n. g. h. can you can that assign centre
00:13:00
stan ford is also building is it's all know they have it's own a. i. n. e. artificial
00:13:05
intelligence in medical imaging and they are more uh also oriented into building an open source
00:13:12
image annotated image trip was it to such as imagine at for radio much
Recommended talks
TEN's pitch
Sept. 16, 2019 · 3:50 p.m.
Clinical Natural Language Processing
Carsten Eickhoff, Assistant professor of medical and computer science at Brown University
Aug. 20, 2019 · 11:05 a.m.
299 views