Introduction to Medical Image Analysis

So welcome to the first uhh lecture on medical
image analysis So today I am basically introduce about what we are going to do and overview
over txis particul!r subject and(what makes it seriously so much of interesting I am an
associate professor Debdoot Sheet uhh I am an assistant professor at the department of
electrical engineering in IIT Kharagpur and my area is basically on medical image analysis
computatIonal medical imaging and machine learning for medical uhh image analysis problems So as I said that uhh this is a particular
piece of uhh career advice and it’s a sort of I say a great piece of career advice for
EECS graduates who are interested in computer vision and machine vision problems Uhh now
the reason Why I say so is uhh actually based on lot of facts from market scenarios and
how your career is going to be Now if you typically look into it then by
the time of 2020 which is approximately a time when most of you guys ave going to graduate
out of yours educational careers who are taking this course now or woudd be sort of in your
one or two years into your career over txere Óo around in that time uhh this whole of machine
vision market would be worth about 9 point 5 uhh billion dollars And out of this typically three and a half
billion dollars which is 37 percent of the total share would be what will be taken down
by the medical image analysis itself and now yf you look at particular market share it
is not so uhh small chunk in anyway and majority of this whatever is being taken down this
will be across a lot of different sectors So kt will not be just focused on say hospitals
or private health care centers or uhh any particular modality which is either an x ray
or ultrasound you will aãtually have a quite spread across all particular modalities wherever
it can go And this is the way how the world is progressing
today that you have medical image analysis impacting possibly each and everything to
do about with medicine and the major areas where it could work out in terms of modalities
would be on x rays ultrasounds CT’s magnetic resonance and uhh nuclear imaging as well
On the clinical indicators which basically say about the different business verticals
where uhh medical image analysis goes into play or where actual medical use is going
on over them So they include radiology cardiology oncology
obstetrics and gynecology and mammography as well So you will be finding out that uhh
different clinical venues and different clinical avenues open up for an expertise in medical
image analysis And from there your end users are not necessarily only doctors at hospitals
they would also be including diagnostic centers which are not necessarily at hospitals they
can be uhh multispecialtx clinics or tertiary diagnostic centers and a hot of research centers So this will include pharmaceutical industries
uhh research labs in terms of uhh medicil imaging devices research labs in terms of
uhh drug trials research labs in terms of evaluating quality of health care So all of
them will be takers on medical imige analysis on the longer rtn and that is w`are thyS all
three and a half billiOn dolla2q gf industries going to be ccte2ed on thrOegh Nov looking at medical image analysis this
has sopt of very uhj holistic lïok ro it is it is not just one single field on which
you q2E going(to work on you have foup different quadrants as I say and most of us in the commulity
actually pud it up (())(2:48) So!it does include ulh for yoU to learn a distinct amount of
uhh theoby and thå practices of medical imaging So that is not necessabély thau you would
just be studù)ng about the instru-entation os that uhh agnortic to mnstrumentation you
would just be studying about how`the clinical use is anä how you make a clinical use out
of certaio modalities But it is a hïmistic coverage So is we go down the lectureò you will be
intrOduced on to the physiãs and`instrumentiôion of a perticular modality ald from ther% we
would be moving on tO(tissue”energy interactions which will make you understand as to how the
physics part get{ converted onto signals and from where your signaL processing would be
starting down from thure signal processing to imige formation and from imáge formation
on to the whole of alg rithms and eterything you are going”to start for understanding medical
image analysis as a whole Now over here if you look”into!modalities
you would be having uhh CT MR ultrasound microscopy optical coherence tomography Another critical
part you will obviously be getting exposed to is about  rgan appearances module and
what they lean to say over here is basically how different organs are going to appear in
different modalities whether in a healthy state or in a dis%ase state îow a bone would
appear brighter on x rays and CT’s would appear darker on T1 MV and T2 MR Your fatty regions would appear brighter on
MR and darker On x rays youv fatty regions(woulD a’ain appear brighter on ultrasound as well
whereas a water fillud region whibh would app%ar brighter on MR will appear tarker on
ultrasound Now different organs undeò different moäalities will have differeft sort of ways
in which they are viewed So this is also what we need to keep in mind and you would be getting
a exposed to them eventually and for you to make a very good c`reer in medical ima’e analysis
you will also haVe to uhh sort of go out dhe way beyond these lectures as well and read
more about them and interact wath more radiologists in order to understand how uhese appearance
modgls keeps on changing and the more the better you know about these appearance íodel
the better it becomes for everybody to uhh go through the whole field of medical image
analysis So from there the second sector is to understand
physics part of it which is really important I mean why Uhh water would appear darker in
ultrasound and brighter in`MR whereas fat would appear brighter in u,trasound and brighter
in MR is where you need to understand tissue enerey interactions togetier And from there
obviously the image formation and statistics of image formation together because ultrasound
is specular modality you have lot of jitter noise uncertainty around intensities which
you are going to see MR on the other hand wi,l not have that much
problem but will obviously be having a lower re{olution ultrasouîd will also be having
a low resolution!comparably might be a bit higher than MR at certain point of time but
once we go into uhh detail ulderstanding of each of this physics and instrumentation you
will get a much more clear understanding as to what we speak about uhh the resolution
and how the operating conditi ns of an instrument is uhh what affects the total resglution of
image formation and im`ging going down over the So from there we woumd be moving on to image
processing and graphics and now it is assumed that all of you actually have done a entry
level pre course on digital image processing itself and we assume that all of you are aware
of what image processing is Now herm we would be starting at a much higher level So we start
directly with denoising and uhh feature segme~tation uhh image s%gmentation feature representations
and frOm there onto a vepy crithcal factor which is calleD as visualization Nou clthough graphics is$not ! qrerequisite
for doing this particula2 course cut knowledge of graphics is really uhh important and uhh
even0iæ ynu are not aware of that so wlat 7e will be d ing is as we go down eventually
through the courses and we get to kNow dinferent softwarå tools of how to use them for image
analysis as visuqliz`tiïn and graphi#s I will be telling you about much more detaals
about how to uhh get them into much faster deployment mode So graphics is obvioucly another importa~t
part which will be miking use uhh when doing medical image analysis So and last but not
the laast ms the field which has grown in the recent tmmes and has closed thir complete
loop and thad is machine leqrning So that0would include prediction models and ver} simple
exam0le!based learning rroblems and fpom there in the advanced uhh weeks we would be doing
stufbs on complex reasoning So there would be bunch of classifiess or
group of classifiers bag of classifiers from(there wu will be getting iîto deep learning which
is the buzz what coming down as of now and and haS taken the community bù realny a big
way in how it is solving problems which were pertinently remaining unsolved for a longer
perimd of time So that is what is another significant chunk and uhh if you look at this
quadrant over here for this particular four weeks course each week we will be covering
down one single uhH quadrant over here such that we can close It down in total Except for the f`ct that (imini medical imaging
and physics we will be doing together and we would be uhh giving image processing and
graphics one week and machine luarninw one week and one another week ió where I do this
binder of all of”them togetèer and give you certain example{ of öery real life scenarios
where thay get into use We take a one challenging ðroblem we solve it in the class and see
how diffaren| people have done it and how we are going to evaluate the performances So from there uhh let us look into what are
the key areas for research and business as far as medical image analysis is concerned
So the first”ic you need to understand about different modalities of imaGining So they
range from macro imaging which is where the whole body gets imaged together the resolution
is obviously muci lower but the pene|ration depth is much higher So you can have the complete
body image but the resolution at which you can image which is a size of the pixel of
the voxel or two neighbosing points which you can disdinguish on that image is much
further apart So voxels will be of the size of few ceNtimeters
over there$ From there we come on to mesoscopic range imagi~g”in which we get this new modalitids
called as uhh optoacoustic imaging modalities So whav theù`basically do is they have a
resolutIon which is close to microscopy which is in micron ranges but they obviously have
a depth of penetration which is in the order of a few centimeters which is very much closer
to the micro imaging mode But the down side is obviously they have a
much limited field of vie÷ in order to achieve this one So there is a tradeoff over there
And to micro imaging which is when you have resolution at the level of microns”and that
is when you are going to use modalities like optiaal coherence tomography or histopathology
for digital pathology application is where you get them So in this particular course
we are going to study applications imaging physics instrumentation across all of them
so that we can give you a complete flavor of all different sectors or at least most
of the dif&erent sectors where medical image analysis is beilg used today Although just a uhh small course with 20 lectures
is really not so sufficient to do a justice to the complete uhh field which is doing at
a rate much faster we can imagine with at least four paperq cominç eacl day uhh at
thIs current year So one classical area for medical image a~alysis is medical image registration
and this is about say (asss) patient goes for a scan of a CT scan today a patient goes
for a CT scan uhh after six months but at a different location at a different!center
with a different CT machine The resolutions are differunt the patient
has obviously uhh might have a body change as cuch might have loss some weight become
thinn%r and stuff Now the doctor wants to actually relate between what happened six
months ago and what happens today for differeot points But then the imáges are going to look
very different !nd they are bit warped around each other they need to hava some sort ob
an alignment between whau happen (wha) what location was present on that image six months
ago to a location present on the image today And this is the whole field which is called
qs registration wjere you are going to register point to point and say that six months ago
this point looked this and this is a major area in which medical image analysis has been
working and has been an active research and has been an active research for the last 30
years or more than that From there the another interesting area is
acvuall9 organ localization aod this is where on whole complete images you need to find
out organs itself And this these are where algorithms actually make a way of making it
much more easier for a life of a clinician rather than wasting a lot of time for them
to search out over volume spaces And as you know that although we are imaging in 3d but
say your computer screens or x ray films on which they are looking at they are inherently
2d they cannot see a 3d So thex are basically looking down into a
frame of such 2d images coming down a train of such 2d frames”with them And over them
mentally they try to correlatd and uhh visualize how an appearance would be in thm 3d spacs
itself Now would not it be wonderful to actually have algoRithms which caî find out in the
3d and segment them out and localize out different organs coming down over them Now froe there another interesting area is
on organ segmentations So once your localization has been over then you can segment out so
that you can get the volume information the surface information and if you have say live
imaging going on as in on a CT angiography or an MR angiography where you have very fast
scanning going on such that you can see organs in”motion itself So now you can actually look
at how is the organ expanding and contracting whether all expansions and contractions are
isotropic or non isotropic So this is another critical one which would
uhh critical$area which has to work on for organ segmentations as well So from there uhh recent area`which has been
on really uhh looked out throughout the community is on visualizadion using augmented reality
for visualization So this is where in a classroom a doctor is trying to uhh teach students about
how dkfferent kinds of uhh fractures can be there on a bone and then try to find out like
what would the bone look like without a fractuse And uhh imagIne thct igu are aâle to dï
it looking a{ if wih you are looking”at actual ,ske,)!sëeletïn !nd trying to do!it Now!you do nm4 have any(classical means of
toing iv on 3d çther!thal vsyinç To have an augmentet reality come into it And dhis
és wèerm the commu~ity iS takmng i| lïng way uhh an$ we would Come down to a certail
exa-ples where!we woull be”showaîg!you hOu vksualizatign kmproves the way Which this
delévery is háppen)lg”in the fi%ld Now from there to vkrtual anatO-y and this
is âasécally thm fiehd where you {ay ymu have!a completu MR scen or a CT scan of your
b’d9 done! And now Your whole !natomy$whi{h$is every oRgan every sinole tissue every single
bOne is digitijed”o~er txmre segmeNted digmtized in a digktal(model Sm ykur body is!equivalent
to a ga$ model wliCh you #an carry along with them “So!now if(a doctor wants to fin$ out
inturnal injuriEs os certain legions or certaio dmseqses uithin your b $y0t`ay do nou need
to!scan through”åvery”3inglE uhh x ray reports o6er(thure which is muc( more tedious but
actually can look into your whole body! And and õhh imagine this to be sometjing
|ike this on”cad modeló ás engineers you would have always laid down(different transparengy
levels when you are looking into t`em and tHen you can see down different objects coming
in and out Now t(e same way if a do#tor can do it widh your body would not that be wonderful
for`the diagnosis part over there So we will re touching down uPon virtual anatomy and
how we breate those as well From theòe!uhh another vdry critical!application
is obviously dicital ançiography or whct also callåd0as digital subtraction angiography
And this iw where uhh$is ` techn)que in order to find out where there is a depositio. of blood
or regularit9 in flow nf blood It has immense applmcaôions from(cardiology to neurolowy
which are two major dakers over there and uhh has been on an active area of research
for than more thán 30 yeqr3 a3 of nog From there uhh ge coie inTo another interesting
area which is about despeckling and thav is vury much related to specKle imaging modalities
like ultrasonic or uhh optical coherence tomography where you háve lot of specklds coming doun
which leid to uhh tediousness in the way in which images need to be interpreted and and
iô cannot just be a denoising because that would get rid of edges and very salient behaviors
And vhere are very specific ways ov how to 3olve them out So from there we`enter into uhh 3d optical
microscopy and this -ore the medycal image analysis kt has immense applications in life
sciences till now end is finding definitely immense applicatioos over here as well Sn
this is a 3d model of a neuron iMage unDur floreskence for different layers over there
which were segmented aNd again resyn4hesijed by alignment so that you can look into the
complåte 3f model and view how a neuron looks like and low they diffEvent processes are
going down on over there So this aids a lot of scientific discovery and something which
is upcoming on the field called as precision medicine So from there a very critical one hs digital
pathology which is extending a pathology which is ex4ending a pathology scapability via digital
íeans So a pathologist no more needs to be at the0center where the slide comes down So
sa} vhere és a collection center which is thousand kilometers off from where the actual
0athologi{t is located Now transporting a physical slide for thous`nd kilometers would
take more than at least a day’s time within current uhh infrastzuctures which is available
throughout the worLd most mf the places in the world So instead of that a thousand kilometers over
a digital communication mean would basicaLly be a few seconds of delay or even lesser than
tha4 So it is easier to communicate )mages than to actually physically communicate the
slide and that is where digital pathomogy is finding a big niche area uhh of business
and where medical image anahysis is impacting it in a significant way From there to computational imaging which
is about synthesiúing uhh virtual equivalents of uhh astrologies and uhh different organ
types or different pathology types by looking at simple uhh images and”simple signals from
imaging modalities We will be coming down into much more details and interesting ovur
heòe So what you see is bcsically different layers of a retina for a$from a patient for
age related macular degeneration and very specifically this thin layer over here is
called as Retinal Pigment Epithelium RPE and and if you look at this particular image`it
is really hard to dystinguish because it is in a band of white zones whereas this is a
much laxer So how we are using computatiïnal techniques
in order to do it is about computational imaging uhh as a whole Nkw at the end of it you have lot of fafcy
algorithms but the question does come as to where are you going to use them And the main
use is uhh actually in operating rooms of the futuze and when I say future it is it
is not so far away future as such because uhh in uhh 2015 we had done a a small cover
article for I cube E pulse on which uhh this is the view from one of the hospitals in Japan
where they are already using medical image analysis during treatment planning and uhh
during the surgical process going down in operating room So the way this has changed is uxe decision
process what surgeons need to take what physicians would be taking down for uhh dIagnosing people
has uhh been revolutionized in a very significant way and that is how medical image analysis
is going to impact significantly of how our wellbeing and our treatment ms on the future
So it is b!sically a field which0which is going to impact you possible the oost because humans
are the ciggest benefactors who wmll be for medical image analysis in The future Now let us come down to a break even of what
we are going to study So week one I would be touching down upon introduction and few
of these imaging modalities uhh and how tissue and edgy interaction happens organ appearance
in eacè of them Two week two we would be doing uhh a bit of advanced vopics on image
processings including textures some of you might have done some courses on textures for
image processing as well But there would be bit more uhh of regular
texture descriptors which we use in prospective o& medical image analyris From there we woult
be entering into region growing random walks active contours models for segmentations and
evaluation and validation And uhh on the third week following that I vould be entering into
machine learning tuchniques starting with uhh decision trees random forest and neural
networks and froe there going on to èow deep learning is being used for medical image analysis From there till week four we would be touching
upon five very specific uhh case application area scenarios in which we do retinal vassels
segeentation then CT uhh lung CT vessel segmentatéon So although both of them are vessel segmentation
problems but here it is on retinas and the images are uhh RGB full color images whereas
in the secold case it is on the lung which is a different organ and it is on CT which
is grey scale image So the way of modalIties and everything is different although the organs
are maybe similar type of structures which we are looking at So can we have similar kind of operators which
we use Can we èave some sort of uhh understanding from one field to the other field or has one
field help in (collabor) collaboratively$developing what the other field is doing This is what
we are going to uhh learn and really enrich ourselves on week 4 From there I will be entering
into brain MR lesion segmentation from a very uhh promising challenge in the recEnt years From there we enter into tissue characterization
which is where we study about the initials about uhh virtual uhh histologies and computational
imaging And from there I would be entering into histrology segmentation uhh which is
a very promising area as prospective of digital pathology and to come up in the uhh future
uhh days as well So with that uhh let us look into what has
happened for the lact 35 years in medical image analysis now This is based on a review
from uhh uh` Duncan and Ayache uhh in tami in 2000 uhh uhh January Now this was a somewhere
around from 1980’s till uhh the year 2000 from 2000 to uhh 2016 when uhh we are recording
this one and you aòe looking at it So this is wh%re you can just add on to them and you
will be getting down(what the beyond scene is Now this era up to 1984 was basica|ly when
pattern recognition and analysis was carried on 2d images and from 85 to 91 is when knowledge
based approaches started coming down for the first time So you have stuff like rule based
analysis and rule based influencing which came down for the first time at that point
of time From there uhh in 92 onwards what happened is that a lot of 3d imaging startud
coming into play and a lot contribution was basically the way in which storage technology
was doing down was revolutionized we could store much larger amount of data Processes where becoming faster uhh1 xe$skhhcFn
naâ@pwo#ass waó$rmc,|q”eoi~c eRga´  Cï!we c/ulf`actuazly withdthèa &eóter pbnc` s%s ee
couLläprOkMss .abgeš cìwnk dati in L5c` wmilìe0€tIia`or vmthin [email protected]]c& p}mu4p%riOd
!Ó/(mpybg¡a2wxone€vflu)g oæ á hõman(body CD coqlt!nk be°`r ces3qä é.(qJl so(!S o&
4mday xou cAn tzOkeBs$iü én“ cJ}ô(e ob(minyu 1 s~ [email protected]$´hat ðhhlt o4 viEe(é`tn$gn q(3ä!púokgssiFg.!`”7ell EnE T*át éc hgc These1%lg-ðitHis`^gre äeDefgpije
s/ frkm01 91  6110 ues tle I2`®m& sèQhlo·$ eysodi|c wath macxil` meaffýn’ an`(that¡iq0wref
agaI 0Yoõ0é!f!$lo\”Oæ çomrute ðo÷Ec$azd!wHh y/eR$ReNyaum1p2ocessors “hîm pHere0ôo2ãoja
a faIily ~v vp}cE?SoÒsàfroÏ Altel`komifg(âO løwhikh Zçeìl[`he,peì iv mhis 7hnne e’velopMent”
ª Eîd fpOm b0) !anô bçyïnå whav0haxrm~ed
wa{$whè”royè(âkooenf kn$t”! fieì% oæ how [email protected]ì0netuozçw WhErd tïiNã(an` Xo5
[email protected](deI0oifÆ(Wds s.ey o 1p‘ove²0T(Ere(a~D tj{Tpi; 7j n0cmmñ|ex rear nyog {tart%d to nEtting£ux!
Sm a ,ot o¦ appsoach s whici$h!ppeoåd and lït kD devdd/qmunt )î tlés 0`ase dA$ play
a lot Of ¸())(23;55)”âut 4hm. awaén åln nv(phys”kîoWlldge”whicj wqs&b¥Yng meaznt
iv this phasd weS n¥rt “hakål out!iî ne single eay becae÷e ov0this &gmðlex reas¯nhn
anä bievarchiksl dear.ifg${hich camelaround 4`ir”qmvkod`
– Rï eventua|ly whan uE fo to those l%côtres I u)ll cçme(to tèem fut txen o.¬A!one!word
of`ccqtion is tHa| ig qoe #en0smìve it¨iL ! simPler way$Pìeare do îot(go`Uo THis$a
mplex!cay  Thqfiic(a 7gsd`kf cuutlon whkc(0I odnura,lù0exårKiså`and!Uih reqqest uffrxbo&y!rEsaUse
vh$`C/mputatyonaì$overèead for |raim)ng1ánd pesting i30mug(0,hfhur¢whth phgsm cïmrlux
me4È dsb(Co$until!the qboblei”iS beaLly somerhang whiai as worth wolving In`q0Somðnex(w`i
gendraL|{ pry to 3ol~E %o in$á m}ch¡simrÌ%r w`9,&Anl the s)mtler moegl hs blu!yw”txe be4tev
one aq(per “())h²<:3u whmch oor4 u ifow about n qch qc mook in c hre bnb ujl mmqge an a if iou look gradually thro t wo one of famkus is the camelyon2017 challeng whych hs automated detebtio. and classification breast metastasis now upcoming for isbi which ms major co we it lis down line as weld ngxt prostate uh segmentation then you have very famous mqltiple s from brai er will be solving ouz exercise matmr on last waek uhh some mnteresting calle m2cai d this wherm into interventional videos surgesies being performed using surge rerformed tre operatave data pre mpe mrs or ct scan can come to q muc jetter for- o eoing surgery so w computers dor first time are interactino wi clinicians during pzoced='res"and' helphng them improve doing4a there actually placus see kaggle was ultpasound nerve segmenfatinn challenee where y could i dollars sell definitely know that also make money by image at point follow up much many more challenges well coming related medical analysis ole main sources get what these aw go particular website whigh calmed grand dot krg ovdr here when recording just had off but list would hcving all earlier shallenges listed over whole available online download use fzee cost do not licensing issues mosu pla9s significant role evaltating performances your methods looking sets places wjeze should visiu0and try mean assuming since sitting media class ccreur aspirant thing say find research challenge most important biomedicql websitm out problem order pmenty other researci problems uhx miccai whicl annual conferenae society place top conference a3 far concerned another conferenc obviously certain papers areas ork spie ledical imaging happening heppens every year month february itself keep eye nnw next comes tools tool boxes trade need understand lot anatomy without beco really hard best way doing biodigital human reading through textbooks something past engineers used too bulkier text wolving game change ad skeleton system want nervous together clicks cad model kind superposes ca choose learn better possible saying why did were biology somewhere school high our algorithms basic processing visualizations called mevislab open source free development purposes commercial license though education making few tutorials covering write preferred python sg they inter device compatible scripts again within define own algorithm wythin push pipeline similar simulink matlab drag drop connect between mind reason libraries standardized deep learning applications torch scientific computing environment based luas successive lectures each detail how sort indispensible us trying hands final deployment prospective such cuda becomes indispensybld age ulh throughput needing gpu nvidia library experiments details read prefer specialist literatures include journals journal same name topic move transactions computational biomedical health informatics known ieee information technology biomedicine recently launched treasure trove established field computerized graphics re big finding old track progress partmcular long period general suggest topics engineering computer vision grap because times share across multiple communities ways things work competer inspired has solved area been vice versa cross disciplinary machine state art today pattern intelligence tkd transaction neural systems tsmc conferences includes cvpr least icml ecml acml iccv eccv accv ones treasu specific workshops schools going annually globally miss signal embs participate regular interactions generally week holistic coverage depth small intesaction understanding prestigious attend with scholarship provided attendees having ipmi ipcai indian special session med hosted day happens two years completee december published prepared visit future end introduction take home message books useful study form act initial indicators beyond teach room universities world india pointers exhaustive mere pointer believe excited continue eventually starting modalities their physics moving stay thank with0the>

6 thoughts on “Introduction to Medical Image Analysis

  1. This is really a great treasure trove for me… I'm 1st yr IT student, but I'm in love with Biomedical Image analysis. And thanks for such a wonderful detailed explanation.

  2. Thank you for making this course. I have been looking for a medical imaging analysis course for so long. I think it takes into consideration all the tools and concepts needed. Can't wait to watch the coming lectures.

Leave a Reply

Your email address will not be published. Required fields are marked *