Big Data and AI in Hepatology with Julius Chapiro

Transcript

• 00:13Welcome back, this session is
• 00:15being recorded. Thank you.
• 00:18So I'd like to introduce another
• 00:20colleague of mine from radiology.
• 00:22Our next speaker is Julia Shapiro,
• 00:23who is an assistant professor in the
• 00:25department, and he is going to talk to
• 00:27us about big data and AI in Hepatology,
• 00:29Julius. Thank you very much and
• 00:34just share my screen and go.
• 00:39I'm OK never going to see my
• 00:41slides and hear me properly.
• 00:43That's great, all right?
• 00:45Well, thank you very much to
• 00:47everyone for organizing that,
• 00:48and I'm very proud to be speaking
• 00:50here at the 75th anniversary of
• 00:53the Yale Hepatology Department.
• 00:55And my topic is to speak about big
• 00:57data in the eye and hepatology.
• 00:58And I want to talk about future of clinical
• 01:01integration of these tools and liver cancer.
• 01:03And at first I would like to say
• 01:06that it behooves us to think about.
• 01:08Medicine and hepatology in particular,
• 01:11uh, from a perspective of data,
• 01:13so this was a projection of Stanford
• 01:16medicine that focused on predicting
• 01:18the growth in healthcare data
• 01:20between the years 2013 and 2020,
• 01:22and they predicted the 15 fold growth,
• 01:24and in fact they were off because in
• 01:26reality the growth was up to 20 fold.
• 01:28In terms of the amount of data that
• 01:30we now have to deal with and on,
• 01:32the right is a typical characteristics
• 01:34of in a large cohort of patients
• 01:36that we once published and.
• 01:38The Journal of Clinical Gastroenterology
• 01:40and Hepatology in 2017 and the
• 01:42table is very confusing and offers
• 01:43a lot of information,
• 01:44so primarily the cancer care
• 01:46can be a data problem,
• 01:48so we do have challenges in modern imaging
• 01:50and intervention that we need to address,
• 01:52and we've been thinking about this
• 01:54issue together with our biomedical
• 01:55engineers and partners in the
• 01:57biomedical engineering sciences,
• 01:59and we have proposed a couple of solutions,
• 02:01so I'd like to talk a little bit about
• 02:05HCC as a data problem so HCC patients.
• 02:08To undergo a multi modality
• 02:10imaging and patient journey,
• 02:12we started with ultrasound screening
• 02:13and then we acquire multiple sessions
• 02:16of multiparametric imaging along
• 02:18the journey of both diagnosis,
• 02:20local regional therapy and systemic therapy,
• 02:22as well as multiple follow-up Mris and
• 02:24every time we do acquire this data,
• 02:26we do get a lot of different parameters
• 02:28that we include into our consideration,
• 02:31yet they are not very well integrated
• 02:33into both the diagnostic therapeutic
• 02:35algorithm but also not very much utilized
• 02:37in terms of how we see those patients.
• 02:39Along the evolving guidelines for therapy.
• 02:42So here are the guidelines that exist
• 02:45for the diagnosis of HCC and Jeff.
• 02:47Wonder have already spoke about Laura's,
• 02:49but I want to say that if there is
• 02:5025 of something across the globe
• 02:52then probably the conclusion is
• 02:53that none of them are really good.
• 02:55So in reality they are unclear,
• 02:58varying and inconsistent definitions
• 02:59of imaging features of HCC.
• 03:01Most references for all guidelines are
• 03:03essentially based on retrospective
• 03:05data and more importantly varying or
• 03:06contradicting or simply no recommendations
• 03:08exist for the assessment of tumor response.
• 03:10You look regional and now also systemic
• 03:13therapy has more agents come into play.
• 03:16We do have a variety of different local
• 03:19regional therapy options and David
• 03:20Meadow have already described some of it.
• 03:22We do have a variety of
• 03:25possibilities to cook, boil,
• 03:26freeze and electrocute the same tumor,
• 03:28and that is essentially dealer's choice,
• 03:31not really based on data.
• 03:32The same thing is true for local
• 03:35regional therapy via the artery.
• 03:36We do have the flavor of conventional taste.
• 03:39We do have the depth
• 03:40taste and we do have wine.
• 03:41Sandy, so very similar therapies with
• 03:43slightly different mechanisms of action
• 03:45and the data around them sometimes is
• 03:47not clear as to what to support for
• 03:50which particular individualized patient,
• 03:52and now a much bigger challenge are the
• 03:54newly introduced systemic therapy agents,
• 03:56of which just a few years
• 03:58ago there was only one,
• 03:59and now we're looking at a large
• 04:01combination first and second line therapies,
• 04:02all with a variety of different mechanisms,
• 04:05and all these are causing different imaging
• 04:07appearances of humor response and the
• 04:09same thing is true with for all these.
• 04:11Regions there imaging challenges.
• 04:13Therefore we need to think about
• 04:15how to transform the burden into
• 04:17value of all this data.
• 04:19On the one hand,
• 04:20we have the data that it's coming to us
• 04:22from the electronic health care records,
• 04:23the genomic sequencing,
• 04:24and we heard a lot of information
• 04:27about that and we now have novel
• 04:29tools like natural language processing
• 04:31and image analysis,
• 04:32and they need computational power
• 04:34at low cost.
• 04:35So then there is this term that
• 04:36is called artificial intelligence
• 04:37that some people use.
• 04:38Essentially biomedical engineers
• 04:40speak of machine.
• 04:41Learning technology that recognizes
• 04:43trends and data and can evaluate patterns
• 04:47based on simply algorithmic learning.
• 04:51And then we have this subgroup of.
• 04:54That is called deep learning,
• 04:56which is based primarily on neural
• 04:58networks that adopt based on data
• 05:00without pre segmentation and without
• 05:03data specific data annotation in some
• 05:05cases and all these things can result in.
• 05:08Hopefully workflow efficiency increases
• 05:10with improved diagnostic accuracy
• 05:12and they can enable us to probably
• 05:15provide us with different practice
• 05:17of predictive recommendations that
• 05:18are better for precision medicine
• 05:20and image guidance.
• 05:20But is that all really true?
• 05:22And it can that really all be applied well?
• 05:24I mean,
• 05:25we've started this this road and
• 05:28essentially mapped out a journey
• 05:30that reaches from diagnosis towards
• 05:31therapeutic decision making
• 05:33interprocedural guidance and follow
• 05:34up imaging and we have worked
• 05:37tirelessly with in partnership
• 05:38with our biomedical engineers and
• 05:40computer scientists on focusing
• 05:42on automation of diagnosis and
• 05:44creating novel imaging biomarkers.
• 05:46Focusing on outcome prediction
• 05:48and therapeutic decision making.
• 05:49Following up for target coverage in the
• 05:52procedure room and then on tumor response.
• 05:54Assessment and the decision
• 05:56for reintervention.
• 05:57Now the vision for all this applied
• 06:00to the multi parametric data that
• 06:02is both clinical and imaging
• 06:03is to achieve fully automated,
• 06:06fast,
• 06:06reproducible and reliable tumor detection.
• 06:08Tissue characterization,
• 06:09clinical decision support system,
• 06:11ultimately providing a tumor board
• 06:13with a probability map that tells you
• 06:15everything you want to know about this
• 06:17tumor and the future decision making.
• 06:19Some of the steps in the nascent
• 06:21steps that we've published are
• 06:22focusing again on live rats and one
• 06:24of the things that we focused on was.
• 06:25Automation of tumor detection and we
• 06:29published this particular work
• 06:31in abdominal radiology in 2021,
• 06:33and the key point here was that was
• 06:35that the convolutional neural network
• 06:36has a high level of performance
• 06:38and automatically delineating
• 06:39liver and focal liver lesions,
• 06:41and it can flag cases with poor
• 06:45performance and therefore functioning
• 06:47effectively as an internal quality control.
• 06:50Another paper that we published
• 06:52focused on the automation of
• 06:55diagnosis and classification.
• 06:56This was a work done with a lot
• 06:58of partners across the board and
• 07:00here we demonstrated the deep
• 07:02learning is able to really classify
• 07:05a variety of different lesions in
• 07:07a millisecond of time per lesion,
• 07:09and this could theoretically help us
• 07:12make clinical workflows more efficient.
• 07:14And here we demonstrated and
• 07:16cross referenced this work with
• 07:18two experienced body trained
• 07:20radiologists and demonstrated that
• 07:22the sensitivity and specificity of
• 07:24the convolutional neural network.
• 07:26Based on algorithm was essentially higher
• 07:29than both radiologists for recognizing a
• 07:32benign from malignant and specifically
• 07:34diagnosing various liberal lesions.
• 07:36Then we also focused on Iraq,
• 07:39path validation and classification.
• 07:40That's a core issue of the work.
• 07:42Here.
• 07:43We demonstrated the deep learning assisted
• 07:45differentiation of pathologically proven,
• 07:47atypical and typical pedicellata
• 07:49carcinoma lesions is possible on MRI,
• 07:52and we demonstrated that
• 07:54even Lara's 4 lesions,
• 07:55namely those that cannot be diagnosed with.
• 07:57Simple imaging criteria according to
• 07:59layouts can be ultimately diagnosed
• 08:02with the help of machine learning and
• 08:04that can potentially help us redirect
• 08:06the need for biopsy and another layer
• 08:08of work that we have added in another
• 08:11piece of information is very early
• 08:13attempt to introduce explain ability,
• 08:16namely to take artificial intelligence
• 08:18as we call it out of the black box
• 08:20and to provide it interpretable.
• 08:21Deep learning system that explains why
• 08:23it came to a specific decision and that
• 08:26I think is very important and that.
• 08:28Point will remain with us.
• 08:30It's an unmet need to explain the
• 08:32decision making of the machine
• 08:34learning algorithm and how could
• 08:35this work in a clinical work file.
• 08:37Well,
• 08:38in reality radiologists could connect
• 08:39to the clinical pack system via VPN
• 08:42and connect it to a research server
• 08:44which we have here in collaboration
• 08:46with Visage Radiologists would
• 08:48then open a multiparametric MRI
• 08:50and initiate the auto automated
• 08:51liver tumor segmentation as we
• 08:53demonstrated that it's possible and
• 08:55with one click classify a lesion as
• 08:57some malignant or benign and then.
• 08:59Uh,
• 08:59the AI would provide a Lawrence
• 09:02report characterizing this entity
• 09:04as specific type of lyrics,
• 09:06and then the radiologists could
• 09:08decline and accept verbiage of the
• 09:10report based on specific features
• 09:12that are highlighted here and and
• 09:14colleague of Mine Joe Cavallo was
• 09:16spearheading this effort with the visage.
• 09:19But then again,
• 09:20the next layer of questions that
• 09:22need to be answered is can neural
• 09:24networks actually help us predict
• 09:26response to certain therapy?
• 09:27And this is something that we
• 09:28thought about and we tested it.
• 09:30Early on in 2017 and a small cohort
• 09:33of patients that underwent the
• 09:35local regional therapy with taste,
• 09:37we put in to the data set.
• 09:39Both age and gender other tumor
• 09:42characteristics as well as imaging
• 09:44appearances and ran a relatively
• 09:46straightforward machine learning algorithm
• 09:48and defined taste response versus taste.
• 09:50Non response as the output
• 09:53classifier and ultimately with the
• 09:54goal to have a good predictor.
• 09:56And we published that model that
• 09:58focused on logistic regression.
• 10:00Random forest in egg PIR
• 10:02and we weren't too hopeful,
• 10:03although the results were quite good
• 10:06in terms of predictability with their
• 10:08Roc curve predictability of 78% and accuracy,
• 10:11we weren't really too hopeful
• 10:12that it's going to take off,
• 10:14but above and beyond expectations
• 10:16were met with this work being
• 10:19cited over 90 times since 2019,
• 10:21where it fully appeared on PUB
• 10:23Med and that really tells you how
• 10:25much this topic is of interest
• 10:26right now in liver disease.
• 10:28And it is definitely worth investing.
• 10:30So how do we see in the future
• 10:33the implementation of these
• 10:34workflows into the clinical value?
• 10:37So right now we have a referring physician,
• 10:39the hepatologist here for example,
• 10:40who send us a patient for image acquisition
• 10:43and discussion on the tumor board.
• 10:44And then we meet in the tumor board,
• 10:46discussed the patients,
• 10:48need clinical and nonclinical,
• 10:49and report.
• 10:50Read the report,
• 10:52demonstrate the images and make the
• 10:53decision on the intervention in the future.
• 10:55I think of the vision can be to have the
• 10:58clinical information be fed into an image.
• 11:01Post processing algorithm that
• 11:02uses those developed machine
• 11:03learning tools to provide those
• 11:05quantitative imaging biomarkers.
• 11:07It's a decision support
• 11:09tool for radiologists and,
• 11:11quite frankly also pathologists
• 11:13and clinicians to really to really
• 11:15help us make the right decision
• 11:17for a specific patient in a
• 11:19more individualized fashion,
• 11:20and that is really true also for the
• 11:22variety of information that we now have
• 11:25through the updated VCP staging system.
• 11:27Now,
• 11:27how do you actually go about
• 11:30machine learning?
• 11:31And I would like to sort of
• 11:33draw a couple of conclusions,
• 11:35slides and say where we need
• 11:36to go so when we talk about the
• 11:38application of machine learning
• 11:39to clinical and imaging data,
• 11:41it's important that AI can't
• 11:42solve every problem in medicine.
• 11:44Not every investigator has a
• 11:45good product or a good idea.
• 11:47We cannot create solutions
• 11:48in the absence of problems,
• 11:49and we also should start with
• 11:51simple issues frequently encounter
• 11:52pathologies with high incident
• 11:54trades rather than very rare finding.
• 11:56So HTC is a really good.
• 11:57Problem and has a lot of good
• 12:00information available,
• 12:00so that's a really good prompt to start with,
• 12:02and I think that we need to focus on
• 12:04the blatant and critical issues in
• 12:06HC rather than the latent and aspirational.
• 12:09And also I would like to suggest
• 12:11that we avoid the black box of
• 12:14actually development and value the
• 12:16Co development between the clinical.
• 12:19Researchers between the physicians
• 12:20and the engineers and computer
• 12:22scientists in order to bring an
• 12:24idea towards a prototype and then
• 12:26validate it clinically and then
• 12:27come up with a
• 12:28product. And this is an iterative
• 12:30process and that is very important.
• 12:32And ultimately I think it remains to
• 12:35be seen how the FDA will handle that.
• 12:37I mean, there is a lot
• 12:38of conversation going on.
• 12:39This is a pipeline that was proposed
• 12:41by the FDA in 2019 and there have
• 12:44been amendments ever since about
• 12:45how a commercially available
• 12:47product is going to be distributed.
• 12:49And used in a real world and the
• 12:51reality of the fact is that that
• 12:54they currently offer voluntary
• 12:56precertification and that all of
• 12:58these tools actually need to be
• 12:59frozen in terms of their algorithm,
• 13:01because before they get approved
• 13:03and once we update the algorithm,
• 13:05we essentially have to repeat the
• 13:07entire approval process and which
• 13:08is very cumbersome right now.
• 13:10But maybe helpful.
• 13:11Sorry to interrupt,
• 13:12but we need you to wrap up very
• 13:14soon as I am about to wrap up.
• 13:16So basically,
• 13:16basically where do we stand for now?
• 13:19And we're just getting started.
• 13:20This is the 8020 rule for many events.
• 13:23Roughly 80% of the effects come
• 13:25from 20% of the causes.
• 13:26This is the parade of rule.
• 13:28This is not the end,
• 13:28but rather the end of the beginning.
• 13:3080% of the effort is yet to be
• 13:32completed and data is too complex to
• 13:34be handled manually and the age of
• 13:36artificial intelligence is here with the eye,
• 13:39we can reach a new level
• 13:40of personalized medicine,
• 13:41and with that I'd like to conclude
• 13:43and say the final thought is
• 13:44really to curb the AI enthusiasm.
• 13:46I think we really need to 1st
• 13:48stand back and focus on the data.
• 13:50And try to build institutional and
• 13:53super institutional databases.
• 13:54That will help us develop those
• 13:56algorithms and validate them.
• 13:58With that,
• 13:58I'd like to thank the Yale Biomedical
• 14:01Engineering and Biomedical imaging community.
• 14:03All of these people are amazing
• 14:04mentors and we have a unique
• 14:06environment to really thrive in
• 14:08this component of science.
• 14:09And I would like also to thank
• 14:11several other mentors that I have had
• 14:13that enabled me to do this kind of
• 14:15research within our wonderful lab,
• 14:16the clinical lab of interventional
• 14:18oncology here at the Department of Radiology.