ai.bythebay.io: Daniel Golden Interview
Recording: ai.bythebay.io: Daniel Golden Interview
you [Music] you hey I is kind of an interesting term and what the the definition that I heard recently that I thought was interesting was AI is anything that's like a few years away from where we are now like maybe five years away so right now this is a conference focused on machine learning applications with deep learning especially although but not everything is deep warning but whether it actually constitutes a I like what people are presenting right now is debatable a lot of a lot of people would say it's really just kind of an implied machine learning model and those have been around for a while the difference is that the machine learning models much more sophisticated now and tend to be more accurate and better able to solve problems that before were intractable so I would find it a little bit of a stretch to for example say that what I'm working on right now is AI since it seems when you get into the code a little bit more mundane than that and the AI seems like it's maybe a few years away but once we get a few years from now we'll have much more sophisticated models I'm sure that are more accurate and able to perform in different situations and at that point we might again say I I as a few years away from where we are now and really we're working on more sophisticated machine learning right now but sort of more broadly i would say AI is a technology that can perhaps reason on its own autonomously without human intervention maybe after a training phase and then also it can sort of interact with a human in a way that suggests some sort of high-level reasoning and empathy for example which I think we're pretty far from with our algorithms right now so the point at which we're able to say that what we're working on right now or a product that we've deployed is really a I I think it's still maybe a decade or so away before we can hit those milestones and so I can talk about what I'm working on right now so i'm working at artists or we have a team that's working on machine learning models automate certain aspects of Radiology post-processing that is when the radiologists review the studies as opposed to when they required that can speed up the radiologist workflow so for example at artists we make a post processing software suite that's cloud-based and is used for post-processing of cardiac MRI so that's mr is done of the heart and one component of that post processing is to contour the ventricles of the heart which is typically done or the past has been done manually in order to get an estimate of the volumes of the ventricles and by getting the volumes you can estimate the vault the volume maximum dilation in maximum contraction and use that to determine how much blood is pumped out of the heart with each cardiac cycle and therefore determine whether the patient is heart failure which is defined in those measurement terms as when the patient is not pumping out enough blood with each cardiac cycle to meet the body's needs and so traditionally this process of doing the contouring has required a lot of manual effort and manually drawing or maybe using semi automated tools that require a lot of correction and so what we bring to the table is a deep warning based system that has been trained on thousands of clinically annotated studies and can now do that contouring automatically so the component of the workflow that used to take 30 or 60 minutes is now not quite completely free because the radiologist has to check it but it's down in order of magnitude from where it used to be so it's really kind of an automation tool and to me I have seen fundamentally what it consists of and it's really a machine learning application I wouldn't really call it AI but to an observer who just sees that sort of they upload a study and magically something that used to take an hour has been completed for them they may perceive that as AI because it has done something that previously took some intellectual effort of theirs automatically I can speak to my field which is medical and a really there's an interesting differentiation between for example autonomous vehicles and and medicine and AI in both those fields and one interesting thing is the population that is potentially at risk for being displaced by those technologies and we the radiologist is that population for medicine and we don't say that we intend to replace the radiologist we want to work with the radiologist and enhance their workflow and make it simpler autonomous vehicles may be there their marketing tools not to say they're going to replace for example tack drivers but certainly for both of these fields you can sort of connect the dots and see that might be if not tomorrow then maybe 10 years down the road that could be the scenario so one thing that we're very careful about in our current work is emphasizing that our goal right now is not to replace the radiologist our goal is to help the radiologist and especially to help the radiologist perform the tasks that take them a long time right now and they're really not intellectually stimulating at all so for example this process of contouring the ventricles of the heart without a visual aid it may be a little bit hard to figure out what that means but essentially you see sort of a bright region in a dark region as drawing the line between the bright in the dark region over and over again on many different slices at different time points on different structures of heart and the task is basically always the same and it's it's a boring task honestly that radiologists don't want to do so the way we're successful is by emphasizing that we're not replacing the radiologist we're replacing this task that you hate doing and by doing so we are speeding up your workflow and making the workflow sort of more pleasant to do because you can spend more time in the more intellectually stimulating activities such as diagnosed in kind of confusing cases or complex cases and working on piecing together this information to make a diagnosis so at least in the medical field I think AI is going to be really successful when we are helping doctors and complimenting the work they do and and removing components that they don't want to do and it's going to be much less successful at least in the near term for things like automating diagnosis where you're sort of claiming I can do what the doctor does better we don't need a doctor anymore and we don't we don't want to go in that direction as a company right now and I think that's going to be less successful because in the end we're selling to doctors and to hospitals and you can't sell to somebody something that you say is going to replace them so for that and other reasons I think taking the doctors full job is not going to be successful but automating the boring parts of the job will be and I think probably my feel there's a medical fields and I don't know as much about other fields but i think that trend will probably be true for a long time another field as well sure so I would say again the key focus is in automating the components of radiology that are tedious boring and what I didn't mention before is also those that have a lot of integrator variations so for example segmenting the the ventricles of the heart even though it's kind of a straightforward task different people do with different ways and as a result as a lot of integrator variation you make a different diagnosis depending on which position is annotating your study so one thing that one advantage we bring us that we could by automating it there's a very consistent way of doing a segmentation and you don't have this variation with different clinicians but in terms of maybe I missed the first part of the question it was remind me to question one more time key focus what what each was doing with my control okay okay so the key focus again is automating the tedious parts of radiology to the clinicians don't want to do what keeps me up at night is something that the the FDA and I'm sure other regulatory bodies and other fields are working to minimize but it's the possibility that a patient will be scanned for example against like the cardiac segmentation thing that is where the patient is extremely different from any patient we train the model on and tested the model on and it's not possible to to test or validate your model and every possible patients games every possible scanner and so on so this is going to happen and the worry is that clinicians by because our our model works so well and so much of the time for example they may become complacent and sort of assume that it always works and sort of stopped double-checking the model and then in this scenario where it does not work as well because it's a very uncommon case they will miss that opportunity to double check the model so I think we really need a way without being heavy-handed about it to encourage the clinicians to double check their work those are the work of the model and make sure that they're always really in the loop both sort of literally because they have to click the button to accept it but also in the spirit of being in the loop they really are cognitively involved in invalidating the models results and sort of an elite the example of where that can go tragically wrong as the Tesla incident where the Tesla autopilot was involved in an accident and the reason is that the driver became totally complacent because ninety nine percent of the time it is it always works you're perfectly are well enough they became complacent and assume that it would work under ten at a time and we need to avoid that local autonomous vehicles and in medicine and really in any scenario where there's our lives at risk for human safety is at risk I thought that on a really nice thing about this conference which is not really as much of an academic conference but is more of an overview of Industry it's sort of the diversity of the different presentations so we've had presentations on natural language processing and visual question-answering on synthesizing music for example using AI on how to manage distributed systems that are performing inference with deep warning models and so on and I think that diversity of applications that are relevant to different degrees but always always interesting to to what we're working on as a commercial company is really interesting and helpful and unfortunately I do have a flight to catch tonight so I won't be attending the rest of the week but I where I I'd be I was employed to more the same coming up coming in the recipe ok Fred you