Player is loading...
AI and Radiology
Dr Amine Korchi
Monday, Feb. 18, 2019 · 11:05 a.m. · 37m 38s
Topics covered: - simple definition AI and Radiology - radiology needs and problems - why we've been talking so much about AI and radiology for the past few years - Can radiologist be replaced by algorithms ? - the radiologist's value chain and clinical workflow - the AI and Radiology industrial ecosystem - AI solutions approved by medical regulators in 2019 - clinical use cases to explore - possible future scenarios. Biography: Medical Doctor, Swiss Board Certified Radiologist & Neuroradiologist and HealthTech Expert fully committed to bridge between Medicine, Science, Technology and Business, bringing Innovation to Market.Clinical AI expert.
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
who will come and go she viewed all your clothes because
00:00:06
you've uh or talk to missus m. u. of
00:00:11
quite interesting profile because we'll reach with the technologies
00:00:16
are you are you will use yeas uh_huh
00:00:20
uh_huh mm someone with your last two this week uh actually use people to really
00:00:28
read read you know real mm e. varies with people for the version
00:00:35
e. the rooms with finer mm so a valuable so uh_huh
00:00:43
mm research yep uh another in a module
00:00:47
university virtues of him right right
00:00:53
so do all of them thank you thank you very much of also up so i'm uh
00:01:02
i mean i'm a medical doctor specialise in medical imaging and the and the radiology
00:01:07
i'm also a help technology consultant mainly for stalked the corporates
00:01:11
and the early stage investment funds uh today i
00:01:14
would like to talk about it i just generic term artificial intelligence but more specifically a deep yearning
00:01:20
and medical imaging so as a party scene comes to pack to sing part time radiology uh the main question i
00:01:27
asked myself is will a i replace radiologists that's the
00:01:31
question i'm always constantly asking about so first let
00:01:35
me just introduce um at the high level though on in
00:01:39
the topic of our our discussion i know we
00:01:42
have here are high level expert in a i sold news are very superficial definition just to define our
00:01:48
our framework a. i has been defined in nineteen fifty five as a as
00:01:53
making intelligent machines that have the ability to achieve goals like humans then
00:01:59
in nineteen fifty five a subset of a. i. which is machine learning
00:02:03
has been defined as the computer's ability to learn without being explicitly
00:02:08
program that was in nineteen fifty nine and more recently since the eighties and
00:02:14
specifically here at age up a a new subset of a i
00:02:18
has been developed name deep learning with multiple layers of neurons deep learning
00:02:24
and especially convolutional neural networks has been very effective in computer vision
00:02:30
imagine recognition and classification and it has an impact
00:02:34
now uh in image and speech recognition
00:02:37
any l. p. natural language processing especially also in
00:02:40
medical imaging robotics and self driving cars
00:02:45
so what about really orgy and really or just which is
00:02:49
a medical specialities all radiologist if you ask google
00:02:53
are defined and that's for sure why if you put
00:02:56
radiologist are radiologist are the smartest parasites reach
00:03:00
over paid reducing the pain of uncertainty they're exposed to radiation happy on common in the
00:03:06
man this is for to what who work tells you about the radiologist i do not agree
00:03:11
with this i prefer the american college of radiology a definition so radiologist our medical doctors
00:03:18
that uses medical imaging technology to diagnose and treat disease it includes ultrasound
00:03:25
its rays computer typography magnetic resonance imaging and was it whole emission come up with
00:03:32
so what is important to to note here is that radiologist
00:03:37
company it's at least two teen years of train
00:03:41
i have don't approximately fifteen years of training including medical school
00:03:45
radiology and said special edition which record fellowships i have
00:03:49
performed to fellowships in your will and muscular skeletal with research so it's a very long training for one person
00:03:56
one person who will work between fifty to eighty to ninety hours a week so that's why
00:04:06
i believe and many other uh medical doctors believe that
00:04:09
radiologist needs help but but why do you need
00:04:12
help so first they need help because there is a a radiologist shortage all around the world
00:04:18
for example if we take the united kingdom with their unified health system the national
00:04:23
health system n. h. s. last year they had two hundred thousand chest it's
00:04:27
rates as a backlog it means that during the twelve prior moment there has been
00:04:34
two hundred thousand just it's right that has not been read by radiologist
00:04:39
yeah waiting in the pack system and it means that some people
00:04:43
some patients has cancer has infections but no one asked
00:04:48
taken the time to interpret these images so it gives you just the sense of
00:04:52
uh how real the shortage he's there is also rising use of diagnostic medical imaging beep
00:04:59
because it is accurate because it uh it gives us the ability to see inside
00:05:03
the bodies without opening and it is also to a data explosion the big explosion we
00:05:10
can explain it by the rising use of medical imaging by number of patients
00:05:14
but also the the the higher technology uh advancement in medical imaging
00:05:20
for example a computer typography city of the thorax and abdomen
00:05:24
could count approximately fifty slice of images each slice gives
00:05:29
an image the transverse an image of the body
00:05:31
i would say ten years ago now we're up to two thousand one thousand two thousand slide because we have a
00:05:37
higher special resolution million metric resolution so would it brings us a lot of data a lot of pixels
00:05:46
computer has has been tough to to to eight radiologist uh we use
00:05:51
sometimes what we called c. a. d. computer aided diagnosis and mammography
00:05:57
which is the uh it's rave to detect breast cancer has one
00:06:01
of the first application of computer aided diagnosis it's a software
00:06:05
the first where rule based expert system would be point in
00:06:09
the image the suspicions area so computer aided diagnosis
00:06:14
are here since nineteen ninety five ninety eight but unfortunately
00:06:20
with the two decades of experience the c. d.
00:06:23
increase the sensitivity to match it to me is there is a lot of false positive
00:06:28
it pinpoints many area off the image that are normal
00:06:33
and and it and it would use the efficiency of the radiologist
00:06:36
and also it's slowed work for all and there is not
00:06:40
a huge adoption of these solution based on rule based expert system
00:06:46
situations are burned out and depressed this is not a just a false prophecy
00:06:51
forty seven percent of radiologist a report sign of burnout or depression
00:06:58
and it not only radios but all doctors uh there is a
00:07:03
very interesting phenomenon which is the medical errors so studies
00:07:08
have found consistently between three and five percent error rate per radiologist per year
00:07:16
so don't get me wrong we're not talking about negligence were talking just about a constant
00:07:22
rate of missed finding for example per human radiologist it gives you a an inside
00:07:28
uh about the limitation of the human brain some eyes
00:07:32
is just an intrinsic limitation of the human radiologist
00:07:41
oh
00:07:46
yeah
00:07:51
so it means that if we're if you take individually one radiologist over one year period
00:07:58
uh out of one hundred cases he when miss three
00:08:03
findings when is three chew marks for example
00:08:10
uh_huh definitely it could have been seen by another radiologist
00:08:14
but this other radiologist will miss three other
00:08:19
the information was into radiology is related to the limitation of the detection
00:08:24
this is all a system and also the cognition of the of the doctor is the physicians cited
00:08:32
finally to tour out of the word population don't have access to medical imaging it means five billion people on or
00:08:39
don't have access to the minimal it's right ultrasound couple of medical image which is quite a lot of people
00:08:46
so for all these reasons radiology needs help computer computer science
00:08:52
is the preferred options because radiologist my essence a pig cells imagines so it could
00:08:59
be quantified it could be a computerised and this doctor doctor remote week
00:09:05
uh it's a doctor or a radiologist in ohio he first introduced the term computer
00:09:10
in the diagnosis in this communication computer aided angles in radiology research plan
00:09:16
and at that time he stated that there is scarcely any repetitive function
00:09:21
in which the computer cannot be of help to us in radiology
00:09:25
and that the application of computer may result in making our radiology department more comfortable
00:09:31
to work in and more productive and this was published in nineteen sixty six
00:09:38
more than fifty years ago and in the same period last week has been the first
00:09:44
to publish the first application of we can call it a high by converting
00:09:50
medical it's race here a chest x. ray from visual
00:09:55
images to numerical sequences on the punch cock
00:10:00
and then he could he have been able to process this punch cards
00:10:04
to look for features of chew marks here you can see in the
00:10:08
core ties to more behind the heart with a level of fruit
00:10:12
here it's the the few we'd hear the what are the x. and that was in nineteen sixty three more than fifty years ago
00:10:20
so why don't we hear so much about the i. today in every journal there is something about it i even
00:10:27
harvard business review in in its professional transition section published
00:10:32
a paper stating that technology will replace many doctors
00:10:37
so first because health care has been digitised
00:10:42
since more than a decade thanks to electronic health record radiology information system
00:10:50
max picture are shooting a communication system by a
00:10:53
pathology information system all dues helped i.
00:10:57
t. infrastructure were put in place during the last twenty twenty five years and
00:11:03
now it's widespread i kiss in the western world and it's growing so we'd bring
00:11:08
us a lot of data that we can process that we can finalise
00:11:12
unfortunately and this is one of the main bottleneck of
00:11:15
a i in medical imaging data are still role
00:11:18
first that's silo they are in proper terry are uh units such as
00:11:24
the electronic health record the packs and they are not interoperable
00:11:30
and they are not structure they are not on any mice and they are not appropriately annotated
00:11:35
so those are they thought they are very important to us but they
00:11:39
are not prepared and it's really hard to prepare the tall
00:11:43
even flown one specific use case but the television
00:11:47
is one of the up you are of what we're seeing today in term of a i development
00:11:53
the advent of very sophisticated deep learning abort especially convolutional no
00:11:57
network and you are the expert here at yep
00:12:02
a a higher processing power with the g. p. u. things the beginning industry
00:12:09
the open source community
00:12:12
the open source community they have has fuelled innovation and i think again hearing a
00:12:17
job we so um the birth of porsche which now we spy talk
00:12:23
big corporations are invasive uh investing massively in
00:12:28
healthcare there's a drawing interest from
00:12:31
microsoft google apple and muscle and all this i would say giant
00:12:38
are trying what they call to disrupt the tree trillion dollar industry that's
00:12:42
how we call it the tree trillion dollar industry healthcare industry
00:12:47
the entrepreneurial scene is free will in actively the field of a i in medical imaging and
00:12:52
you will see the number of startup working on specific use cases and the national
00:12:58
a i plans such as france china and the us for all these reasons
00:13:04
a i is under the spotlight and especially for health care especially for half get so back
00:13:10
to uh the main this use case of i i. which is a computer vision
00:13:16
image classification object detection here is the error rate on
00:13:20
imagine at visual recognition challenge and you see uh
00:13:25
very at a tremendous uh mm i think increase in accuracy
00:13:31
the error rate was around twenty five percent in two thousand eleven and it
00:13:34
it's either human level in two thousand fifteen less than five percent
00:13:39
the other thing i liken this chart is that the first improvement were performed
00:13:45
by university of toronto new york but then it's corporation will microsoft
00:13:52
and that to add jobs just hinton i i know i think you know him one of the
00:13:57
key opinion either in a i to compare radiologist too wide a coyote in the cartoon
00:14:04
geoffrey hinton state that people should stop training radiologist now it was i think around two thousand fifteen
00:14:10
or sixteen which is a clear cut conclusion and i can understand him because he's from the
00:14:17
i would say the injury engineering aspect non clinical aspect is not on the on the
00:14:22
floor in the hospitals but he's in the research uh computer science lab a i
00:14:27
and from his perspective if an a. i. algorithm is able to detect
00:14:32
a cat or a dog or that in an image why
00:14:36
with me being able to detect the to work on the brain an right or an infection
00:14:41
and pneumonia on it just it's way id and those are both images with excels
00:14:48
for this specific use case stan ford has a performer study the checks necked algorithm
00:14:55
which allow the detection of pneumonia which is the infection off the long box
00:15:01
in chest it's way so chest it's right it's this kind of images in this one more that it's
00:15:07
we can explore just by just it's right or white c. t. even by m. r. i. checks it's right
00:15:12
is the first step it's the cheapest it's a two d. images it's widespread using the checks net algorithm
00:15:20
this is the activation not you can detect the pneumonia here behind the heart the accuracy each see that
00:15:27
the accuracy of four radiologist including special lies radiologist in
00:15:33
the chest image they're doing only just image
00:15:36
so what's interesting here not only the accuracy was better
00:15:42
but the algorithm underlies the whole batch of images one
00:15:46
hundred and fifty time faster then schumann radiologist
00:15:50
and this is a critical point especially when you're here that in the n. h. s. they have approximately
00:15:55
two hundred thousand similar just it's way sleeping in their packs no one is able to read them
00:16:03
so andrew engine i think you you know we also he pasta just we after this uh this study
00:16:10
asking should radiologist be worried about their jobs so i would like to answer
00:16:15
andrew n. j. that this study is i would say a great achievement
00:16:19
the great advancement for the signs but this algorithm the checks net detect one
00:16:26
does is pneumonia on one imaging model it it is extremely narrow
00:16:33
it does not replicate what i'm doing as a radiologist what i'm doing is looking at the whole image and
00:16:41
concluding with a full report stating the reasonable abnormality screening over hundreds of abnormality
00:16:48
hundreds off normal variants soul what they have which is is extremely narrow
00:16:54
and this is the main i would say limitation of this algorithm from my clinical perspective
00:17:03
the second thing is that the main publication or research project
00:17:09
currently available in medical imaging and i are in city cool retrospective studies
00:17:15
they are testing algorithm on data set retrospective data
00:17:20
sets testing validation but those are not
00:17:25
a i i'll go written tested on the real world prospective intervention all studies
00:17:32
in the real world in the real because system within the work for all
00:17:36
this is the second critic and limitation of these kind of states but
00:17:41
again if we come back to this algorithm it's very narrow any does not reflect what
00:17:47
the radiologist do during his daily jot and also why the radiologist is paid for
00:17:55
mm insurance does not pay the radiologist exclude pneumonia from just it's right
00:18:00
they pay us to finalise the chesapeake straight as a whole come with a conclusion and then go
00:18:06
forward in patient management so it brings us to what i call the radiologist value chain
00:18:14
the value chain of these kind of doctors so it again when
00:18:17
a non radiologist for example an emergency doctor calls me
00:18:23
and asks me for an exam i need a chance it's right or i need a city for
00:18:28
my patients he has a a an abdominal pain and i think he has an appendicitis
00:18:33
so the first step is discussing with this doctor doing the patient history evaluation
00:18:40
and after that we decide that there really is room for medical images it is appropriate
00:18:46
to make this exam for this patient in this situation and then we have
00:18:50
to select the exam type should we do an ultrasound or c. g.
00:18:54
i i just sound isn't much cheaper than a c. g. i. interest sound isn't is not using x. right
00:19:01
it's rare ionising radiation that it's raised by themselves they can induced cancer
00:19:07
then we have to find the slot on the panic in some countries you
00:19:11
have to wait four to six months to having him right for example
00:19:15
skin is very important then for each patient you have to customise
00:19:20
the exam if someone has a chew on the brain
00:19:24
it is not the same protocol i someone hasn't multiple sclerosis form and right
00:19:29
you have to customise it and during the exam when the patient
00:19:33
is in the machine during the exam acquisition you have to
00:19:35
adapt the protocol in order to have to the best image
00:19:39
the best information to arrive to diagnosis or follow up
00:19:44
then after the exam acquisition when you have the the images you have to control
00:19:49
the quality sometime make some additional images you have to discharge the patient
00:19:54
and many times especially for elective patient outpatient you have to give
00:19:59
them what they are asking doctor have i what have something
00:20:03
it's very hard to to tell them immediately the answer but you can i personally say that the
00:20:07
image were of good quality and i'm gonna send uh the result to his uh twos physician
00:20:14
then after that you have to process the image to have a better quality were to
00:20:18
enrich them to mix between two sequences or also to do some quantitative analysis
00:20:26
and one thing that i like to to mention even if it's not
00:20:30
present in uh the and all the paper written in this field
00:20:33
how we display the images on the screen this is very crucial for work for all because you have
00:20:38
to see the exam made today but you have also to see the exam made three months
00:20:42
before and two years before to compare for for what you have to
00:20:46
display multiple sequences which gives you multiple different view of the brain
00:20:52
and then you are able to better and i like how to display all
00:20:55
these images on your screen by itself is the word for acceleration
00:20:59
and in between is the effectiveness also on the pictures of the rigid that's my point of
00:21:03
view i i could help also here after displaying the images on my screen then
00:21:10
i can look at the image and lies then and make the detection
00:21:15
of the anomaly here there is a protection of those pixels and they try to classify it
00:21:19
which kind of two more benign or malignant and put a diagnosis do school work tasks
00:21:25
are the computer vision tasks the main computer vision task as that which is the perception today in the a.
00:21:31
i. n. medical imaging landscape after detecting this is classifying
00:21:35
diagnosing you have to to produce a full report
00:21:41
and then to communicate the results this radiologist value chain
00:21:48
we are doing this in different setting for screening for example when you have to apply a medical
00:21:53
image exam at the population level for example to
00:21:56
detect breast cancer in women weren't fifty role
00:22:02
for two yeah i emergency rule is it normal abnormal or is it emergent not an urgent
00:22:08
for diagnosis of this is an for monitoring for following up a a cancer after treatment
00:22:14
ah i deep learning convolutional neural network jan
00:22:18
most probably will have a huge effect on detection classification diagnosis by essence
00:22:24
those are the core of the computer vision analysing it says looking for
00:22:28
pattern but i'm convinced that the effect on a i would
00:22:33
be huge and would be faster on what we perform before and
00:22:37
after for an automatic and appropriate exact selection examination selection
00:22:44
uh what the doctor one unknown radiologist doctor asks for an exam for these
00:22:49
patients what we called or the rain there is a high proportion
00:22:54
i don't have the number but there is a high proportion for my
00:22:56
expertise of nice ordering for example he will ask for city
00:23:01
where we should one m. right or we will ask for city one
00:23:05
it makes right is sufficient or not or sell a i can
00:23:09
help streamline this process reduce cost accelerate
00:23:14
work throws same for automatic skinny
00:23:19
protocol examination quality control image processing
00:23:23
structured synoptic reporting i i
00:23:27
can also help us three populate reports with quantitative data extracted for ages
00:23:33
while markers are and we have also to keep in mind that
00:23:39
the work of the human radiologist includes also some highly value tasks that
00:23:44
are very hard to replicate by automation which are multi disciplinary
00:23:48
meetings were multiple doctors surgeons internet it's in radiologist you kermit since it's
00:23:54
in the same rule and speak about one patient sharing taught complementing
00:23:59
the view to arrive to the better treatment and better madam for
00:24:03
this patient image guided intervention where what it's gonna help but
00:24:07
even the most advanced product now in the pipeline which i'm close to this
00:24:12
community of intervention or robot it's a nice human interaction with patient referring
00:24:18
physicians training and research activity will still need human radiologist so that's an
00:24:23
overview of what radiologist human radiologist and the radiology department do in
00:24:30
the work flow on the field today each of these so a radiologist steps
00:24:37
are screwed tonight by the industry and every startup every wishers team
00:24:42
is trying to to find a good use case and to create the
00:24:45
product may in these two steps to detection classification and like
00:24:50
knows it's not only the the the the the the demons g. and finance
00:24:55
we which are the leaders in medical imaging industrial walking on this
00:24:59
but i would say the most active are the tech design such as
00:25:03
microsoft i. b. m. who will a hole in the also some
00:25:09
eastern companies like i abide by do percent and a huge number of
00:25:13
starts the start that are highly active in this field each of
00:25:20
the startup is working on one specific use case this one is swiss
00:25:26
and this one's based insert there also some data science institute
00:25:34
in situ that bridge between hospitals and that a scientist with the goal to develop a product
00:25:41
which will be incorporated into works all of the
00:25:44
partners hospitals and then tested and iterate
00:25:49
and the good the bad example is the n. g. h. and b. w.
00:25:52
h. sent perfectly good that assigns in boston so here's an overview
00:25:58
of the software is using deep yearning for radiology cues case so all
00:26:06
those are starts i i dock in his right eye cards their
00:26:11
brightest red bailouts id x. was the first to obtain an f. d.
00:26:16
a. for column was uppermost screening of they up there but it
00:26:22
retinal patch for retinal scans and regarding the indication or the use case
00:26:28
the most common is in track ran out hemorrhage on city
00:26:32
but mammography has also a lot of traction and many companies are focusing on
00:26:37
mammography because mammography is a screening program each women depending on the country
00:26:44
after forty five or fifty years should have uh one mammogram your early or
00:26:50
wants to your to detect cancer so with at the population level the
00:26:55
number of imaging is very high and i are good pure can help for
00:26:59
tree arch to tell the radiologist that twenty percent of the mammograms
00:27:05
available in its database our priority to see because they might be some findings
00:27:10
here so this is the f. d. a. approval there's also the c.
00:27:14
approval not this begins per presentation what is important to note is is
00:27:19
the what we call that a high cost the castle is like
00:27:23
like a canyon it's a like a a in five infringe sell the canon
00:27:32
is the gap between developing in actuate algorithm with high accuracy ended the
00:27:39
medical news in the real world application x. k. this yeah that i
00:27:45
mean all in the technology life cycle adoption it exactly what we
00:27:51
are uh with this thing today actually a and c. approval are just uh
00:27:59
they're approving the safety and effectiveness they are approving marketing cleans it
00:28:05
means if i submit an application to the f. d. a. stating that
00:28:10
my algorithm is effective at fifty percent is accurate at fifty percent
00:28:15
if i can prove it they will give you the f. d. a. approval but would you like to be tested with a
00:28:21
fifty percent accuracy okay i'm exaggerating let me just to give you
00:28:26
a sense f. d. a. and c. d.'s you just
00:28:29
uh approve of safety and effectiveness but it does not give you
00:28:34
the proof of clinical effectiveness and costs does the new a.
00:28:41
i. device a little better patient and calm at reasonable cost in the real this is the this is the most
00:28:48
important question this is what we call the help technology assessment and the cost effectiveness that if you're a i i'll go
00:28:55
it is able to in improve the income outcome the clinical
00:29:01
outcomes and to reduce the healthcare cost this is a
00:29:05
real of that and then insurance we'll reimburse or product and then
00:29:09
you would create helpful product and the profitable company if
00:29:13
it's a commercial company and finally in order to do is
00:29:18
we have to shift from retrospective you cynical research setting
00:29:22
to prospective intervention or real world clinical trials what we call randomised
00:29:27
control trial where you want the mice people to treatment
00:29:30
a or treatment b. and you compare the outcome with a i without i so let's take the mammography case
00:29:40
so as i explained mammography is an image of the breast
00:29:43
uh it's a widespread screening program in many countries
00:29:48
so i'll z. i. a. have approved in nineteen ninety eight the
00:29:53
computer aided diagnostic software for mammography doesn't rule based expert
00:29:58
says that in the mid twenties less than five percent of
00:30:02
the screening mammograms we're allies with c. d.'s all the
00:30:06
way there was a little uh accept ration adoption of this
00:30:09
technology only if when you drive into the building of
00:30:14
the health care system you will find that insurance where reimbursing
00:30:19
all of these exams but doctor in reality doesn't
00:30:22
rely a lot of them because there is a a lot of false positive because it slowed work for all
00:30:28
and in two thousand fifteen there the this publication off a very well
00:30:34
performed study could be seen uh the journal of underground medical association
00:30:38
uh was comparing the diagnostic accuracy of not working with and
00:30:42
without computer aided detection the conclusion was clear cut
00:30:47
computer aided detection does not improve diagnostic accuracy of mammography insurers pay
00:30:53
more for c. n. e. we know well established benefit
00:30:56
women in servers in the us pay approximately four hundred million
00:31:02
dollar per year only for using this computer aided
00:31:06
detection is reimbursed between seven barber exam to twenty the offer
00:31:10
some private insurers and this is only for mammograms so
00:31:17
i'll just talk to why then that would be the school of thought deep learning and the medical application
00:31:23
of deploying is very effective in computer vision and i
00:31:27
want to display here some other application outside
00:31:30
of radiology so first digital pathology digital pathology has
00:31:36
been lagging behind medical imaging in the television
00:31:39
where now i can say it's with that on all the radiology department are dished out how they are
00:31:46
feeling free with packs and i. r. i. s.
00:31:50
some if not the most digital pathology department
00:31:54
are still working on what we call whole slice imaging so there's slightly teach you
00:31:59
for example when they remove the remove it you more they would slice it
00:32:02
into teams lies put it onto a glass and looking on it
00:32:06
with a microscope they are not the to light it is the trend i believe then in the next five to ten years there
00:32:12
will be the same situation in pathology with a full the television of
00:32:15
the batteries because it it uh it will accelerate work flow uh
00:32:20
it will decrease holes all the burden of having this whole slice images
00:32:25
and by digit digitising these images we can apply computer vision
00:32:31
but that's another aspect interesting indeed the pathology and this aspect is also present
00:32:37
in a medical imaging but there is a local requirements between two
00:32:43
apologies to medical doctors because the the same slide two two different pathologist
00:32:48
and you will find usually difference in the interpretation again here
00:32:54
deep learning who would help to have a more objective more raw produce
00:33:00
able results e. k. g. e. k. g. is uh the electric
00:33:07
a recording of the heart is also a well studied application to
00:33:11
detect heart i read now to predict so them cardiac arrest
00:33:16
or other serious heart problems skin lesions detection of metal
00:33:21
one other skin cancers so we have now
00:33:25
a lot of um well a bit studies showing that the accuracy in prospective settings of
00:33:31
a deep learning algorithm it's yeah it's the dermatologist level of that action of of men and
00:33:36
well this is an inch or inside view of the a ball well the colon
00:33:42
this is also an application of a a i could detect abnormal
00:33:46
because only asian face recognition is also studied for mood
00:33:52
a sec active disease nor nor the united disease and finally retinal
00:33:58
scans and i think you heard about this that is
00:34:00
a ladder by a good where they use retinal scan not
00:34:05
only to detect diabetic written apache that's what ophthalmologist
00:34:09
do but they use this kind of the writing on the backside of uh of the height to
00:34:15
detect cardiovascular risks age if the patient smoke or not
00:34:20
the level of cholesterol and here it's the analysis
00:34:24
by the pruning algorithm of this commonly available images innovative
00:34:29
new by markers and this is also one
00:34:33
of the power of the burning taking some images that
00:34:37
we use for big noses and extracting information about
00:34:39
the patient that we were not able to extract before so in all these application the main added
00:34:45
value of a i today it's accuracy and speed and the price is for automation right it brings us
00:34:53
to the future so i see multiple scenarios i
00:34:59
foresee multiples and i saw the first scenario
00:35:02
the more romantic size is a poor automation
00:35:06
of medical image replacement off radiologist
00:35:10
by soft works the second scenario is deeper learning technology as all
00:35:18
makes medical imaging accessible enough for non
00:35:22
radiologist doctors for example cardiologist or
00:35:25
topic east surgeons to ally themselves the images of their patients now
00:35:32
the knee the radiologist and by allowing these non specialist doctor in medical
00:35:38
imaging to interpret medical images then you take over this activity and
00:35:44
it would be very happy to do this because they will take over one crucial part of their diagnosis and and for what tool
00:35:51
they will have more control on their work flow and more important they will increase the revenue
00:35:57
and this is this is not a false prophecy cardiologist have already taken coroner rubber thing
00:36:05
a coke or geography or some of the heart and now they're taking
00:36:10
over heart and my soul if the burning a decree is
00:36:16
the bar at the um the barrier to entry to medical imaging
00:36:19
and that is they would take it all in both scenarios
00:36:23
the radiologist is out there know infer radiologist i think news do do
00:36:29
scenarios as that are repeated strain uh from my perspective i
00:36:34
see more a combination between man and machine man 'cause machine equation
00:36:39
better than man versus machine i believe that deep learning
00:36:43
we increase the accuracy of radiologist that we make the work
00:36:50
flow more streamline it and they will also a
00:36:54
low radiologist to uh achieve a very deep river of
00:36:58
image analysis renovating new buyer markers it's eating
00:37:02
the limitation of our visual systems and our cognitive system and up more than that i believe that the
00:37:09
burning and technology will most probably increase the affordability and
00:37:14
accessibility of medical image analysis and health care
00:37:19
for the underserved population the five billion people want or so so finally i
00:37:25
believe that i i will disrupt radiology but it won't replace radiologist
Recommended talks
Digital lessons learned from musculoskeletal radiology
Patrick Omoumi
Feb. 1, 2019 · 11:37 a.m.
1878 views
Keynote session: Accessible Bioanalysis for the Developing World and the Point of Care
George M. Whitesides, Harvard University, Cambridge - USA
Oct. 26, 2017 · 5:09 p.m.
230 views