Seeing beyond the scan in neuroimaging

In this episode, we explore the intersection of AI, machine learning, and healthcare through the lens of neuroimaging and epilepsy diagnosis. Dr. Gavin Winston shares insights from his work using MRI data and machine learning to uncover subtle abnormalities in brain function. We discuss the cultural and ethical barriers to AI adoption in medicine, how predictive data analysis could transform the diagnostic workflow, and what the future holds for medical imaging in a world increasingly shaped by intelligent systems. Featuring: Gavin Winston – LinkedIn https://www.linkedin.com/in/gavinwinston/, Website https://deptmed.queensu.ca/people/gavin-winstonChris Benson – Website https://chrisbenson.com/, GitHub https://github.com/chrisbenson, LinkedIn https://www.linkedin.com/in/chrisbenson, X https://x.com/chrisbensonDaniel Whitenack – Website https://www.datadan.io/, GitHub https://github.com/dwhitena, X https://x.com/dwhitenaLinks: Detection of Epileptogenic Focal Cortical Dysplasia Using Graph Neural Networks: A MELD Study https://pubmed.ncbi.nlm.nih.gov/39992650/Machine Learning in Neuroimaging across Disciplines https://research.hkbu.edu.hk/page/detail/22685Automated and Interpretable Detection of Hippocampal Sclerosis in Temporal Lobe Epilepsy: AID-HS https://pubmed.ncbi.nlm.nih.gov/39543853/Literature review and protocol for a prospective multicentre cohort study on multimodal prediction of seizure recurrence after unprovoked first seizure https://pubmed.ncbi.nlm.nih.gov/38582538/Deep learning in neuroimaging of epilepsy https://pubmed.ncbi.nlm.nih.gov/37473486/Non-parametric combination of multimodal MRI for lesion detection in focal epilepsy https://pubmed.ncbi.nlm.nih.gov/34619650/Detection of covert lesions in focal epilepsy using computational analysis of multimodal magnetic resonance imaging data https://pubmed.ncbi.nlm.nih.gov/33567113/

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:03Speaker 00

  Welcome to Practical AI, the podcast that makes artificial intelligence practical, productive, and accessible to all. If you like this show, you will love The Change Log. It's news on Mondays, deep, technical interviews on Wednesdays, and on Fridays, an awesome talk show for your weekend enjoyment. Find us by searching for The Change Log wherever you get your podcasts. Thanks to our partners at fly.io. Launch your AI apps in five minutes or less. Learn how at fly.io.

• 0:44Daniel Whitenack?

  Welcome to another episode of the Practical AI Podcast. This is Daniel Whitenack. I'm CEO at Prediction Guard, and I am joined, as always, by my co-host, Chris Benson, who is a Principal AI Research Engineer at Lockheed Martin. How are you doing, Chris? Hey, I'm doing very well today. How's it going, Daniel? It's going great. I'm excited to kind of switch it up from all of the talk of language models and agents, although maybe that will feature some somewhere in the conversation today, but but turn to something that's really exciting and I'm really happy to that we can feature on the show, which is all about neuroimaging and and machine learning and AI, which will be an

• 1:28Daniel Whitenack?

  exciting topic to learn about. Today, we have with us Gavin Winston, who is Professor at Queens University. Welcome, Gavin.

• 1:35Speaker 02

  Good afternoon. Thank you for the invite, Daniel.

• 1:37Daniel Whitenack?

  Yeah, yeah. Well, it's great to great to have you with us. Maybe if we can kind of backtrack from even the even the machine learning AI side of neuroimaging, even just to kind of the context of neuroimaging, a lot of our audience might be in the technology space or software developers or, you know, business leaders. Could you help us understand a little bit, you know, when when we're talking about neuroimaging, we're talking about some of the work that you're involved with? What does that? What does that mean exactly? How does it impact people's lives and kind of the healthcare system? And, you know, the treatments, etc.?

• 2:18Speaker 02

  So neuroimaging is a pretty broad term. It covers a variety of different techniques. And essentially, the general concept of these approaches is that we're looking at the structure and function of the brain. You can look at other parts of the nervous system as well, but for my work, I'm particularly interested in the brain. So for things like structure, you can look at perhaps, is there an abnormality within the brain that's causing someone's seizures or other types of neurological problems? And on the function, you can look at which parts of the brain are responsible for different functions that people have. For example, which parts of the brain are people using for language, which parts of the brain are people using for memory and other tasks like that. So I guess I'm more concentrating on MRI, but there's a whole variety of techniques,

• 3:03Speaker 02

  things like CT scans and MRI scans and PET scans. There's a whole series of different things. But essentially, they're just techniques to look into the brain from the outside and try and give us an idea of the structure and function, how these might be altered. So I guess the most common thing people would see would be MRI scans as a thing, which is what I mainly work on.

• 3:23Daniel Whitenack?

  And maybe also give a little bit of context for some of the, I guess, more of a historical context for, you know, maybe when and how machine learning or AI techniques started coming into an intersection with neuroimaging or MRI. How recent has that been? Is that something that's been going on for quite some time? And yeah, maybe just give a little bit of context there as well.

• 3:55Speaker 02

  I think maybe we can step back and think about how neuroimaging has developed over many, many decades.

• 4:01Daniel Whitenack?

  That'd be great, yeah.

• 4:02Speaker 02

  So of course, back at the beginning, you first developed the concept of doing X-rays and we could do an X-ray of the skull. But that wasn't particularly helpful because you couldn't see the brain inside the skull. You could just see the skull itself. So then people started developing other pretty barbaric techniques where you would inject air or other things into the brain and then you could somehow highlight them that. That was pretty risky and certainly not that helpful. But I guess it started really coming in the 70s, 1970s, when CT scans started to become available. So you could then get a nice illustration of the brain or other structures from the outside.

• 4:42Speaker 02

  And then that developed further to MRI, which gives us much more detailed pictures looking at, looking at the brain. And we can get higher and higher resolution and much more detailed now.

• 4:55Speaker 02

  As these techniques have developed, of course, we've got more and more data to analyze. And the more and more data we have to analyze, that's when we start thinking, well, how can we use techniques such as machine learning to learn from this vast amount of data we're now starting to collect? So when you think about an MRI scan, you could have a resolution of 256 by 256 by 256 voxels. So it's a three dimensional picture. But then you have multiple different types of MRI scans, looking at different, different types of things within the brain. So you have an absolute ton of data. And when someone, he was a radiologist, in other words, a neurologist, sorry, a physician that specialized in assessment of scans, they have a lot of information to look at.

• 5:41Speaker 02

  And that's extremely time consuming. And the number of scans we're doing is going up and up. So now we're starting to think, well, how can we help and sort of save time and also make it more easy to detect the abnormalities or automate things more, more effectively? Because the vast amount of data we have now, that's, that's not possible for us to now keep, keep up with.

• 6:01Chris Benson?

  I'm curious, with the different types of scans and with all the data that they're producing, and now that you have machine learning techniques available, has that changed some of the choices that doctors are making in terms of what's possible yet? Or is that still more in the future? You know, how, how has machine learning changed the practice of medicine in these areas?

• 6:24Speaker 02

  As with many things, when you're looking at the integration of machine learning and artificial intelligence into medical practice, the uptake of these techniques is, is can be fairly slow. There's a fairly big chasm between what's possible technically versus what is actually used in practice. There's obviously a lot of concerns that people have around data quality, ethics around using the data, the accuracy of any techniques you might be using, because of course it's going to be used for for humans that are undergoing different diagnoses and treatments. So there's a big gap between what can be done and what is actually being done in practice. So the uptake, I mean, there's a lot of potential there,

• 7:09Speaker 02

  and you always see things being developed, but the uptake has been a little bit slower than I would like as someone working in this field. But I think, given what we can do now, if we think forward over what's going to happen, definitely these are going to become much more important over time.

• 7:26Daniel Whitenack?

  And maybe before we get into how like a machine might process some of the data that you're talking about, if we just consider the human, the physician, or whoever it is who's looking at some of the data coming off of these scans, what are what are some of, I know there's probably a whole variety of things that could be discovered in in that data in the imagery, right? But just by way of example, what what might the human observer be looking for in in the neuroimaging data that would give them a sense of, I guess it would be a diagnosis or something to investigate further, or a potential abnormality, or whatever that is, what are some of those things? And, you know, what would they be

• 8:10Daniel Whitenack?

  looking for, you know, with their own with their own human eyes?

• 8:13Speaker 02

  So when you are seeing a patient as a neurologist, you will get the description of the symptoms and examine the patient as well. And that will give you an idea of where in the brain might be involved by whatever is going on. And then depending on the symptoms and the types of symptoms, you can get some idea of what types of abnormality could be occurring in that part of the brain. But the role of the neuroimaging is really to confirm that there is an addamati in that region of the brain and what it is. So what where is it? And what is the what is the problem? Because you may have a list, so-called a differential diagnosis, you have a list of the possibilities it could be based on what you've managed to get from the patient. But until you actually do the scan, you don't know exactly which one it is.

• 8:56Speaker 02

  So when a radiologist is the physician looking at the scan, they will be provided the scan plus the clinical information and some hypothesis that the clinician is thinking about where it might be and what it might be. So then they're obviously going to look closely at those areas and try and identify something that correlates with that. And for a radiologist looking at things, a lot of it is about pattern recognition and recognizing things that they've seen before. So sometimes it's very hard for a person to define what it is in the image that they see that tells them it's a certain thing, but they've seen this before and that's what it is because there's a certain pattern that somehow they've learned that that represents that.

• 9:41Chris Benson?

  Got a quick follow up. You know, as Dan asked the question and you were answering it, it occurred to me how little I know about the topic as a layperson, obviously. And you kind of started with the notion of kind of two things, that there was structure and there was function of the brain. And I'm kind of curious, could you take a second, kind of backing away from the data side of things, but just kind of going into that and talk about, if you're a neurologist, what is the relationship between structure and function in the practice even before you get to the data? So like, how do they relate in terms of diagnosis and your evaluation of a patient?

• 10:23Chris Benson?

  And how might that inform bringing data into it, as Dan brought up in that last question?

• 10:29Speaker 02

  So most of the scans that are done in day to day life would be scans looking solely at the structure of the brain. So for example, if someone has presented with symptoms that you think represent a stroke, we're wanting to do a scan to work out, is there a stroke and where is the stroke? So you're looking at the structure of the brain. So pretty much all the clinical scans out there being done are scans specifically looking at structure. So there's a whole separate side of imaging when we're talking about MRI imaging, which is looking at the function of the brain. And this is used in specialist centers and in specialist situations. So for example, if we're contemplating doing a surgical treatment on the brain to treat

• 11:14Speaker 02

  some underlying condition, of course, we don't want to know just what the brain looks like, we want to know which parts of the brain are performing different functions. We know in general that certain tasks are localized to particular parts of the brain, but each person is individual and it may be slightly changed by the underlying abnormality they have. So there are certain scans you can do to look at the function. So for example, if you want to do some brain surgery near the visual pathways of the brain, you might want to identify where the visual pathways are by some form of functional imaging. Or if you're doing surgery near where language function may be, you want to know exactly where in the brain language function is, so you can try and avoid that area if possible.

• 11:57Speaker 02

  So these are so-called functional scans when you're looking at the function of the brain. But that's far, far less common in day-to-day practice.

• 12:05Daniel Whitenack?

  And this may be a, maybe an interesting question, but it's not often we have someone who's on the show that's both an expert in kind of machine learning, AI type of things and, you know, neuroscience or neuroimaging. I'm wondering, you know, from over the years, of course, we've had many people on the show, there've been many parallels to, you know, between neural networks and the structure of the brain and how these things are, you know, maybe modeled after, after one another. I'm wondering from a kind of expert in the, in the field who's also applying machine learning and AI techniques, just how maybe complicated or, or different the, the, the

• 12:50Daniel Whitenack?

  brain might be than these kind of, you know, neural networks or deep learning systems.

• 12:57Daniel Whitenack?

  That yes, are very powerful, but you know, at least in my understanding at their root contain, you know, very simplistic components and certainly aren't as efficient as the brain in in many ways. I don't know if you have any thoughts on that, but I figured I would, I would take the chance because we don't, we don't often have this, this intersection of expertise on the show. So.

• 13:20Speaker 02

  Yes, it's a great question. And I think if you think about neural networks, as you mentioned, of course, they are based on biology and what happens in, in reality. But there's quite a big difference between what we're simulating and what the reality is. And a lot of it is around the, the scale. So when we do, when we have neural networks, although now of course, we can have much more complicated neural networks with the computational power we have now than we used to, you don't realize just how complicated the brain is, just how many billions of neurons it has and how they're all vastly interconnected. So that's, that type of complexity has been

• 13:59Speaker 02

  very, very difficult to emulate. And even when we try and emulate the nervous system of a very simple organisms that only have a few hundred neurons, it's very difficult to replicate that precisely. So how can we possibly do the same for a structure like the human brain, which has so many more orders of magnitude neurons and synapses there. But yes, it's based on, based on some underlying anatomy and function within the brain, but it's just, there's a big gap between the level of complexity of what we simulate and what we actually are doing in our own brains.

• 14:35Daniel Whitenack?

  So Gavin, now that, now that we have a bit of context about neuroimaging and, and the brain in, in general, I, I'm wondering if you could give us just now kind of zooming in, in a level down, I guess, or more focusing on the, the computational techniques. At, at kind of a high level, how would you categorize the ways in which, you know, machine learning or AI is being applied to tasks related to, to neuroimaging?

• 15:04Speaker 02

  There are a number of different ways that we're using a machine learning applied to neuroimaging, tackling various different parts of a patient's diagnostic and treatment journey. So what example would be trying to classify scans as to what, whether they contain an abnormality or not? So that's a simple classification task. A lot of the literature out there, they collect data on some healthy individuals without the underlying condition. Then they also collect some data on some people with a particular condition. And the aim of the machine learning algorithm is to try and classify whether someone has a particular condition or not on the basis of the imaging.

• 15:47Speaker 02

  You can publish quite a lot of papers doing that, but the question is, how useful is that in the real world? Because it's very unlikely you'll be saying, us wanting to ask it, do you have condition X or not? Which is essentially what the classifier is doing. You have a patient and you want to know what condition they have, not do they have condition X or not? So therefore, you need to go a much, much higher level than that, trying to classify a much different abnormalities. And another way that classification could be used is not just, does a person have condition X or not? But it may be a condition that could be in various different parts of the brain. So what you're trying to classify is which part of the brain is affected by this condition?

• 16:32Speaker 02

  And which part of the brain is normal, healthy tissue? And that's particularly something I do in my work in epilepsy. So epilepsy is a condition with recurrent seizures coming from the brain. And in some cases, those seizures may be coming from a particular part of the brain. And we want to know where the abnormality in the brain is causing, that's causing those seizures is located. Apart from that, what I've just mentioned about diagnosis, then another thing that's being used is, how can we use imaging to help us guide what type of treatment someone should have? Which is the best treatment, given this scan we have? And also, can we infer some information about the prognosis of a patient?

• 17:15Speaker 02

  If we treat them, what is the likelihood of a particular outcomes? So for example, if it's if it's brain tumor, for example, what's the likelihood of someone recovering from that or passing away from the underlying brain tumor? Can we predict that from our imaging data? So it's used for a whole variety of things, from diagnostic purposes through to treatment options and prognosis as well.

• 17:37Chris Benson?

  So I'm curious that the the use cases that you were kind of talking about applying data and machine learning techniques to were really fascinating to me, you know? And could you take as an example, one or or more of them and kind of talk about which ML techniques that maybe our listeners are familiar with, you know, that they've used in different in their own jobs that are unrelated? And talk about, well, I I take this and I apply it to classifying what part of the brain is affected, you know, by epilepsy? And and kind of tie in a little bit about what a practitioner that is listening might have done themselves and they might never have occurred to them to use it in this way?

• 18:22Chris Benson?

  And just kind of, I'm trying to kind of tie the technique with the application itself a little bit. If you could kind of just take us through a couple of examples possibly on that.

• 18:31Speaker 02

  With the first example I gave of just trying to classify whether someone has a disease or not, typically, we extract a number of different parameters from the imaging. And that that could be designed in many different ways. And then we have a known output. So this is they either have the disease or not because we we've decided to scan people predefined disease or not. So this is essentially a supervised learning approach, and you can just use typical techniques such as a support vector machine that you may have seen in many other approaches. This is a very common thing that is used in this type of approach. Then some more of the some more of the advanced imaging techniques when you're looking at the whole three dimensional picture of the image.

• 19:15Speaker 02

  This is very well suited to techniques such as convolutional neural networks. They're extremely good at doing this type of imaging analysis. So I just mentioned two techniques, but these are the probably the most common two things you will see in the literature, but there are plenty of other options as well.

• 19:32Daniel Whitenack?

  And with that, I'm imagining that, you know, I've seen a lot of problems in supervised learning with, you know, I get a data set and, you know, I train the model. I've seen computer vision techniques. I'm imagining there's a number of unique challenges on the technical side with what you're doing, which may be related to the data, maybe it's related to the scale, maybe it's related to the correlations, maybe it's related to the complexity of the number of target classes. Or, yeah, maybe just help us understand kind

• 20:12Daniel Whitenack?

  of where, what are the challenging elements in any of those aspects as you apply those kind of general purpose techniques in this specific, specific case?

• 20:23Speaker 02

  One of the biggest challenges is having access to the data in the first place. So when you look at techniques such as large language models that have a vast trove of data on the internet that they can tap into. Unfortunately, we don't have the same thing when we're approaching neuroimaging. So if we want to train a model for whatever task we're trying to train it for, the more data we have in general, the better we can train the model. But the question is, where do we get this data from? And that's one of the biggest challenges. So if you, if you are running a project at a university and you're scanning subjects for a research study, you may be able to get 100 subjects. But

• 21:04Speaker 02

  100 subjects to someone in machine learning, that is a miniscule number. So that, that is nothing like hundreds of thousands of data points that you might want. So the first thing is getting the data. And there have been approaches to try and combine data from other centers and multi-center approaches and so on, but still were limited by the amount of data. And of course, if you're using a supervised learning approach, the data has to be labeled as well. And depending on the nature of the label, that could be extremely laborious and time consuming, and maybe inaccurate depending on, and who's doing that. If we're just classifying someone's having a condition, a disease or not, that's a pretty easy classification to make. But again, mistakes could be made.

• 21:49Speaker 02

  But if we're doing some of the more complex things such as trying to work out which part of the brain is affected by a particular condition, then often the way that this is done is by someone labeling the image. In other words, someone draws manually in the brain which part of the brain is affected, and that's what's used to train the algorithm. But given how I mentioned that there's so many voxels in the brain, the amount of time it takes someone manually to draw around the abnormality in the brain, that can be extremely time consuming, several hours per subject. So if you have to do that on a hundred subjects, that's a lot of manual work. So it's extremely difficult to get the data

• 22:29Speaker 02

  plus also the labels for the data. So that's a pretty big challenge for any clinically based study using machine learning and AI algorithms, just getting that in the first place.

• 22:42Chris Benson?

  Given the limitations of the available data, in this field, is there any role for synthetic data, you know, in some other areas, you know, unrelated that is acceptable and in others I've heard reasons why it's not? Is any level of synthetic data that you're producing to support the research, is that a possibility? Is that something that you stay away from? Just curious.

• 23:05Speaker 02

  Sometimes people augment their data by synthesizing slightly modified versions of the data they have and use that for training.

• 23:14Chris Benson?

  Right.

• 23:15Speaker 02

  So that is something that can be done. But I don't think you can synthesize data completely from scratch. You can just modify some data, your existing data you already have. Yes, that's certainly a possibility.

• 23:27Daniel Whitenack?

  And, you know, mostly you've talked about some of these approaches, you know, convolutional neural nets, support vector machines. Again, a lot of, I think a lot of the questions we're asking are coming from a place of those not working, not working in the field and questions that people might have, you know, from outside of the medical field. You know, a lot of what I've heard is that there is definitely a, you know, more of a burden for explanation of certain predictions, right? And a sort of audit. Trail potentially of of how decisions were made. Obviously, that's easier potentially with a with a, you know, machine learning model than with a deep, deep neural net.

• 24:11Daniel Whitenack?

  You know, what, what is the reality there in terms of the techniques available to you? Not so much technically, but from, but from a practicality of how they might be, might be applied. I guess one thing is proving something in, you know, in, in a paper, right? But then actual adoption of that could be, could be challenging. So what are the realities there?

• 24:33Speaker 02

  Yeah, we already discussed that unfortunately, implementation is well behind what it would ideally be at this stage. And I think one of the challenges is if you develop an algorithm and you present it to a physician and say, look, this does, does such and such. They want to understand how that's working. They want to be able to trust the algorithm. So if you've developed an algorithm that's meant to detect disease X, how is it making that decision? Is it doing something that makes sense to me? Because ultimately, of course, a lot of the AI algorithms could be, just appear to be like a black box. You don't, you've got your input, you've got your output, but you don't really know what went on between those two steps.

• 25:14Speaker 02

  And particularly for physicians less familiar with these techniques, they want to know what's happening and so they can trust the algorithm. Because at the end of the day, you're going to be making a decision that can affect someone's life on the basis of this information. So you want to be sure how that, how that works. So certainly some of the imaging analysis programs are starting to work more towards the concept of explainable AI, as you mentioned. So actually, I can mention one study here that's, this is a study that I'm contributing data to, but it's led by my colleagues at UCL. It's the MELD study. So multi-center epilepsy lesion detection study.

• 25:57Speaker 02

  And this is a study which is meant to help us detect wherein the brain is causing epileptic seizures. And they're collecting data from many different sites across the world. And one of their key aims is to develop something that explains why the decision is being made. So the output of the algorithm is not only just This is where in the brain we think there may be an abnormality. There is also then an output that says these are the features that were different in that region of the brain that have led us to believe that that is where the abnormality is. And then when we look at that, we could potentially even ask a radiologist to go back and look at that part of the brain again and say, look, this is what we're seeing as potentially different.

• 26:42Speaker 02

  Are you able to now see that on the image?

• 26:45Chris Benson?

  I'm curious, before, you know, a couple of points in the conversation, you've talked about kind of the, the speed of adoption of the technologies and you addressed a little bit about, you know, what some of those root causes are in terms of the desire to understand algorithms such as that on behalf of the practitioner, the medical practitioner. Culturally, you know, within the profession, what do you, what kinds of mind shifts do you think are, are going to need to take place, maybe to accelerate adoption or, or what, you know, what is that natural progression that you're seeing? Because as we're seeing these technologies flood into completely different, you know, industries across the globe and in different capacities, we're seeing these kind of, kind

• 27:30Chris Benson?

  of cultural struggles, you know, within given professions on that. And I'm really curious, as you, as you look at these, at these doctors that are at some level adopting the technology and moving forward with it, recognizing the benefits, but also some of the challenges based on their traditional thinking, what do you think it'll take to get there for that profession?

• 27:50Speaker 02

  Well, the use of AI in our day to day life has now become so widespread. I think people are becoming much more acceptable of the technology as a concept. But when we're working with clinical data, what are the limitations we have is what are the ethical considerations behind that? And that's one barrier to adoption. So where is the data that we've acquired from a patient going? Is it, if we're submitting it to some algorithm, where is it being processed? Where is it being stored? Is it being kept? Is it being used for other things in the future? Is it, is it covered by privacy law, et cetera? So there's a lot of considerations there. But if we can get something which addresses all of, all of those concerns, I think now

• 28:33Speaker 02

  is the time in the next decade and so on, we can actually start to get these things much more widely adopted because there is that, the widespread acceptance of AI now.

• 28:45Daniel Whitenack?

  So Gavin, I'm wondering, obviously you've done a variety of research in this area and are aware of other things that are going on. You know, we've mostly talked about kind of the context, the background of the problem, the technology, the challenges. I'm curious a little bit about the potential impact or performance. So let's say that you're doing, you know, one of these studies, maybe you could give an example. What is the kind of comparison between, you know, a human maybe that's doing this sort of review manually and identifying either certain diagnoses or parts of the brain?

• 29:29Daniel Whitenack?

  Or I imagine, you know, things get more complicated as the problem gets more complicated. But what's kind of the human performance level? And where have people been able to push the kind of machine learning AI performance level? Now, granted, as you mentioned, there's still challenges to overcome with the adoption. But I'm curious, at least in the studies that you've done or have seen, how is that stacking up? And maybe also, what does it seem like these models are really good at? And then maybe what are some of the open challenges that are, you know, not addressed currently?

• 30:08Speaker 02

  Yeah, so the performance of humans in addressing whether there's an abnormality on the scan is very, very variable. There are a lot of studies out there that look at inter-rater performance between different people looking at the same type of data. And unfortunately, the performance in agreement can be quite poor in some cases. And if you look at trying to detect some subtle alteration in the brain, there are studies out there that show that the more expertise and the more highly trained the specialist is, the more likely they are to detect the abnormality. So if you have someone who's a highly specialised neuroradiologist only looking at scans of people with epilepsy,

• 30:50Speaker 02

  they are much more likely to detect it, any abnormality, than someone who's a neuroradiologist looking at all sorts of scans. And they're in turn more likely to detect it than someone who's a general radiologist, not just specifically looking at neuroradiology. So there's a lot of, as with anything in life, the more highly specialised and trained you are for something, the more likely you are to detect things on a scan. Having said that, the reason we want to try and use AI sometimes is to look for things that aren't easily detectable by the human eye. There are certain things that are very hard to visually perceive, but there are patterns in the imaging data there that can be picked up by the algorithms.

• 31:33Speaker 02

  I think that's a very strong case for the use of AI, for things that really are not visually apparent or easy to detect. But we can use it in a lot of other aspects of our life as well, all through the whole process. So if we're putting in requests for scans, of course, there's going to be a waiting list for scans because many scans are being requested. Perhaps there's some way you can look at the information given on the patient and decide which ones are the more urgent, which ones are more likely to yield something abnormal that affects how I treat the patient. And then, once you've done the scans, there will be given, a radiologist will be given a list of scans to look at.

• 32:18Speaker 02

  But if you can pre-assess those scans with some form of algorithm that prioritises the scans and say, "These five scans appear to have an abnormality. Look at these five scans first." That's much more useful. And then you can leave the other hundred scans that are probably normal till later you prioritise the ones that are potentially going to change someone's treatment. And then, I mean, we could go on. There's just many steps in the whole process that you could integrate this into.

• 32:46Chris Benson?

  I'm curious, as you're kind of defining, you know, kind of how it's changed the, the current workflow where, you know, the, the automation is, the model is kind of pre-selecting scans for the radiologists. But yet, you know, with the, with models advancing so fast right now in terms of the capabilities, would you expect that that to kind of remain the relationship between the model's capability and the human radiologist? Or do you think that that's going to evolve over time? And, you know, is that, is that a constant evolution that you anticipate? Or do you think there's some sort of, of kind of human AI, you know, steady state? You know, in terms of the, the collaboration between the technology and the human at this point?

• 33:31Speaker 02

  Yeah, I guess, I mean, we're often asked, is, is AI going to replace the physician?

• 33:36Chris Benson?

  Indeed.

• 33:36Speaker 02

  Essentially that type of question. I, I personally do not think it will. But I think it's going to be a technique that facilitates and helps. In other words, it's going to augment the abilities of whichever type of physician you are. It will make your workflow more efficient, more smooth and so on, but it's never going to completely replace the human aspect. So for example, now when we use, if we use any of these techniques and identify things on imaging, which we think may be relevant from the AI, we always then present it back to, to the team of physicians to look at the scan again and then determine whether we think that's relevant or not. It's never, it's not replacing what, what we do. It's giving us some information which we then go back and review ourself.

• 34:21Speaker 02

  So for example, there's something that may have been overlooked or very hard to see in the first place. An algorithm that says, possibly there's an abnormality here. You can then go back and confirm or refute whether you think that's the case or not. So it's essentially, I think it's going to remain this type of thing augmentation of what you do, but not a replacement.

• 34:41Chris Benson?

  So, and if I could just follow up for a two second thing on that, and I'm not disagreeing with you, but I am curious because people will say, oh, I think the human will stay in the loop and stuff in that. But, but why, what is in, in the way that you're looking at it? Why do you think the human will stay in the loop? And I'm not, I'm not arguing against that or saying that's a bad thing at all. But I've, I've gotten that when I tell people that I think a human will stay in various workflows in other industries and stuff. That's, that's a common question is they go, well, why? Based on what you're telling me, you know, with the model increasing, I'm kind of curious what your, what your kind of foundational belief is there.

• 35:19Speaker 02

  Well, however good algorithms are, they do make mistakes. And one of the big issues in the type of algorithm I'm talking about, which is trying to detect when abnormality in the brain is causing seizures is a lot of false positives. So the technique looks good on, on, on, on the basis of that we are identifying where the abnormality is, but it doesn't address the fact that maybe three or four other brain regions that were also identified that were not involved at all. So you're getting a lot of false positives. So the performance of the algorithms is good, but it's far from perfect. So I think that's a key reason why you're always going to need some human oversight and looking into that.

• 36:03Speaker 02

  And of course, then there's a whole separate issue of what is the legal responsibility? If an algorithm says X and you make a decision based on that, and it turns out what it said was wrong, whose responsibility is that? Is it the person who used the algorithm? Is it the person who wrote the algorithm? Is it a physician? Which physician is it? It's a difficult decision. So I think there's always going to have to be some form of human oversight in the process.

• 36:34Daniel Whitenack?

  And how on that front, on the legal side, are jurisdictions or governing bodies, or whatever the relevant association would be, are those bodies keeping up with this sort of work and kind of ahead and putting the legal frameworks in place? Are they catching up? A combination of the two? What guidance is coming down? And how does that legal situation look as of today?

• 37:06Speaker 02

  That's not something I've looked into a lot because each country has its own different rules. But in general, unfortunately, the legal system lags a lot behind the technological innovations that are occurring in the world. So there's definitely a lot of scope to working out all of these legal issues and how they best dealt with.

• 37:28Daniel Whitenack?

  Yeah. And I know, speaking of things changing in the world, obviously, we've seen a huge shift and perception shift and shift in technology and AI over the last couple of years with GenAI and language models and vision models and all sorts of things. I imagine that there is discussion in the research community around how these types of generative technologies might play a role in, maybe it's the decision support around this type of work or other things. Is there any perspective you have there?

• 38:08Daniel Whitenack?

  Or have you crossed paths with folks that are considering those sorts of techniques in addition to the machine learning, traditional machine learning models or CNNs or deep learning type of models? What's the kind of status there in terms of reception or integration of this latest wave of technology?

• 38:32Speaker 02

  That's not an area I work in much myself, as I mainly concentrate on the image analysis side of things. But I think that goes back to what I mentioned earlier about the potential triage opportunities of these type of approaches. So given this textual information on why we're doing a scan plus access to someone's medical records. What are the likely possibilities? What's the probability of these things actually being the case? Which is the most urgent scan that we should be doing doing next? I think I see it. I see that area as when it's going to be the most useful.

• 39:07Chris Benson?

  As you were looking at this, you know, it's I find it really fascinating. I'm of an age where I'm in my mid 50s and and having lots of medical procedures and and seeing the advancement of technology pretty rapidly. As you look at this area that you focused on so much, where do you envision the technology taking the profession and the tools of the profession? And how do you think that will will go forward in terms of as as we have ever more AI capabilities and algorithmic capabilities and more data available? What does the what does the future of imaging look like to you? And where do you think it might go? And are there any particular things that you would like to see happen as you've as

• 39:52Chris Benson?

  you've thought about your work and where it's going and, you know, the rate of adoption? I'm just curious what that future is.

• 39:59Speaker 02

  Yeah, but what I would like to see is as a radiologist sits down to report the 20 scans they've got, instead of scan one to 20, it's scan most likely to be abnormal to least likely to be abnormal. So you've already got your list in the order you should be looking at them. And when you open the scan, rather than just seeing the scan itself, just just the scan, the there's an algorithm been run on it already, a form, the type of analysis you're interested in, and that's generated a report, some recommendations, some ideas, which has already been fed back into the radiology system. So when you open it, it's not just the scan, it's the scan plus sort of computer generated recommendations report.

• 40:43Speaker 02

  So you already have that information ahead of you so you can then focus on those areas and that and then potentially detect things that may have been overlooked before or get things more quickly. And then, once the report has been done. Then you need some way that that's going to be fed back to the treating physician in a useful manner. At the moment, what happens is the radiologist writes a report, and that is sent electronically, or sadly, in some cases by paper, to the referring physician to look at when they're getting around to looking at it. But maybe there's some AI that can be put in at that stage, which can generate a recommendation or prioritization

• 41:24Speaker 02

  of those information, those results, so that the physician that actually requested the scan in the first place is alerted sooner when there's a significant abnormality that needs to be addressed.

• 41:35Chris Benson?

  Sounds good to me.

• 41:36Daniel Whitenack?

  Yeah, definitely. I think that's a great picture to paint as we, as we close up here. Gavin, it's been a great experience having you on the show. I encourage people in the show notes, we'll, we'll include a couple of links to where you can find some of Gavin's papers and, and presentations. I encourage you to check it out. Really appreciate the work that you're doing, Gavin, and appreciate you taking time to chat with us. It's been great.

• 42:02Speaker 02

  Okay, great. Thank you very much.

• 42:10Speaker 00

  All right. That is our show for this week. If you haven't checked out our Change Log newsletter, head to changelog.com/news. There you'll find 29 reasons, yes, 29 reasons why you should subscribe. I'll tell you reason number 17, you might actually start looking forward to Mondays.

• 42:31Speaker 04

  Sounds like somebody's got a case of the Mondays.

• 42:34Speaker 00

  28 more reasons are waiting for you at changelog.com/news. Thanks again to our partners at fly.io, to Breakmaster Cylinder for the Beats, and to you for listening. That is all for now, but we'll talk to you again next time.