{"version":"1.0.0","segments":[{"startTime":5.345,"endTime":9.75,"body":"So, thanks ever so much, um, for inviting me to present this talk here at"},{"startTime":9.75,"endTime":10.925,"body":"this, this ancient place."},{"startTime":12.265,"endTime":16.015,"body":"As Martin said, I'd like you to tell you about what's happening in this area"},{"startTime":16.015,"endTime":17.765,"body":"of AI as applied to medical imaging."},{"startTime":18.225,"endTime":22.725,"body":"And I'm gonna try and explain my understanding that I've built up as I've done"},{"startTime":22.725,"endTime":23.025,"body":"some work in this area about what's going on under the bonnet of these systems,"},{"startTime":23.025,"endTime":30.525,"body":"um, and some of the challenges in development of these systems."},{"startTime":31.025,"endTime":35.165,"body":"And then move on to a bit of, uh, speculation about what might happen next."},{"startTime":35.705,"endTime":41.045,"body":"So I'm hoping I can encourage all our curiosity and express some of my optimism,"},{"startTime":41.045,"endTime":44.605,"body":"optimism and enthusiasm for, for kind of what's going on."},{"startTime":44.865,"endTime":49.342,"body":"But at the same time, I wanna steer away from the, the hype and feverishness"},{"startTime":49.342,"endTime":52.925,"body":"that is kind of really, uh, prevalent in this, in this area."},{"startTime":53.225,"endTime":58.372,"body":"And as a bit of concrete context for this, there's a guy called Jeffrey Hinton,"},{"startTime":58.372,"endTime":58.715,"body":"who's a, a very renowned, uh, pioneer in AI in the us And at, at"},{"startTime":58.715,"endTime":68.325,"body":"a research meeting in 2016, uh, he said that we should stop training radiologists."},{"startTime":68.385,"endTime":72.244,"body":"And in his words, he said it, it's just completely obvious that within five years"},{"startTime":72.244,"endTime":74.045,"body":"deep learning will do better than radiologists."},{"startTime":74.915,"endTime":79.045,"body":"Well, thankfully, I mean, I was just finishing my radiology training at that stage, \u003claugh\u003e."},{"startTime":79.105,"endTime":83.362,"body":"Um, but thankfully most people didn't listen to 'em 'cause we'd be in a real"},{"startTime":83.362,"endTime":86.485,"body":"mess if we had no new radiologists arriving at the moment."},{"startTime":87.265,"endTime":88.885,"body":"So I'm a radiologist."},{"startTime":88.985,"endTime":91.285,"body":"I'm, I'm a clinical doctor, specialized in imaging."},{"startTime":91.505,"endTime":96.048,"body":"I'm not a computer scientist or a mathematician, but I do want to stray into"},{"startTime":96.048,"endTime":96.351,"body":"that area of, of looking into the clockwork of these systems, partly because I just,"},{"startTime":96.351,"endTime":103.925,"body":"I think it's really interesting and I hope you will too."},{"startTime":104.105,"endTime":107.687,"body":"And, and partly because I think it's, it's quite often glossed over in these types"},{"startTime":107.687,"endTime":108.165,"body":"of talks."},{"startTime":108.625,"endTime":113.925,"body":"And I think a wide appreciation and understanding throughout society of, of how these systems"},{"startTime":113.925,"endTime":116.045,"body":"work is, is gonna be helpful."},{"startTime":117.185,"endTime":119.165,"body":"And like I said, I do think it's fascinating."},{"startTime":119.305,"endTime":123.365,"body":"So you've heard a bit about, about me from, from Martin."},{"startTime":123.525,"endTime":125.885,"body":"I wonder if I could find out a bit about this audience."},{"startTime":126.005,"endTime":130.171,"body":"I dunno if, if you are willing to raise hands if anyone's kind of worked"},{"startTime":130.171,"endTime":131.005,"body":"in healthcare here."},{"startTime":132.675,"endTime":134.685,"body":"There's all this, there's, there's lots."},{"startTime":134.855,"endTime":135.205,"body":"Great."},{"startTime":135.425,"endTime":140.085,"body":"Um, how about kind of mathematicians or, or computer scientists or AI experts?"},{"startTime":140.665,"endTime":143.285,"body":"Oh, right, we've got, we've got quite a few as well."},{"startTime":143.585,"endTime":149.035,"body":"So I hope I don't raise too many of your eyebrows with my explanations about"},{"startTime":149.035,"endTime":150.125,"body":"what's going on."},{"startTime":150.825,"endTime":158.081,"body":"Um, all of us here are healthcare users and all of us therefore generate healthcare"},{"startTime":158.081,"endTime":158.565,"body":"data."},{"startTime":160.175,"endTime":164.312,"body":"The Gresham College lecture that you're listening to right now is giving you knowledge and"},{"startTime":164.312,"endTime":166.795,"body":"insight from one of the world's leading academic experts."},{"startTime":167.695,"endTime":169.195,"body":"Making it takes a lot of time."},{"startTime":169.695,"endTime":172.901,"body":"But because we want to encourage a love of learning, we think it's well worth"},{"startTime":172.901,"endTime":173.115,"body":"it."},{"startTime":174.055,"endTime":177.235,"body":"We never make you pay for lectures, although donations are needed."},{"startTime":177.695,"endTime":178.995,"body":"All we ask in return is this."},{"startTime":179.635,"endTime":181.925,"body":"Send a link to this lecture to someone you think would benefit."},{"startTime":182.265,"endTime":186.181,"body":"And if you haven't already, click the follow or subscribe button from wherever you are"},{"startTime":186.181,"endTime":186.965,"body":"listening right now."},{"startTime":188.025,"endTime":189.645,"body":"Now, let's get back to the lecture."},{"startTime":192.265,"endTime":195.085,"body":"Uh, it, this is an, an area of that affects all of us."},{"startTime":195.265,"endTime":199.285,"body":"And, and about half of us are gonna encounter cancer in our lifetimes."},{"startTime":199.285,"endTime":202.885,"body":"And, and perhaps some of you here have already had that experience."},{"startTime":203.385,"endTime":206.285,"body":"So I, you know, this is, this is important for everyone."},{"startTime":206.365,"endTime":215.788,"body":"I think, um, I'm gonna be reading, largely reading my presentation 'cause I just get"},{"startTime":215.788,"endTime":217.045,"body":"mixed up."},{"startTime":217.045,"endTime":221.954,"body":"If I don't, um, if I end up talking about the next slide without moving"},{"startTime":221.954,"endTime":224.245,"body":"it on, somebody shout slide at me."},{"startTime":224.925,"endTime":225.005,"body":"\u003claugh\u003e,"},{"startTime":226.495,"endTime":230.81,"body":"First of all, before we dive into the kind of AI side of things, I'd"},{"startTime":230.81,"endTime":231.098,"body":"like to give you a whistle stop tour of the incredible and varied technology that"},{"startTime":231.098,"endTime":236.565,"body":"we use in medical imaging."},{"startTime":237.025,"endTime":242.584,"body":"Um, so medical imaging has evolved from experiments producing the first radiographs at the end"},{"startTime":242.584,"endTime":242.955,"body":"of the 19th century to becoming kind of completely ubiquitous and central to the provision"},{"startTime":242.955,"endTime":248.885,"body":"of healthcare."},{"startTime":249.655,"endTime":255.733,"body":"We've, we've got the famous radiograph here of, of, uh, Anna Renton, the the wife"},{"startTime":255.733,"endTime":258.165,"body":"of the X-ray discoverer, Wilhelm Renkin."},{"startTime":259.144,"endTime":263.983,"body":"And these type of plane x-rays, uh, are routinely used in healthcare and, and carry"},{"startTime":263.983,"endTime":266.565,"body":"on being an essential part of practice today."},{"startTime":266.865,"endTime":271.14,"body":"And, and they're fast, they're inexpensive, and very often they're all we need to take"},{"startTime":271.14,"endTime":272.565,"body":"to make the correct diagnosis."},{"startTime":273.345,"endTime":278.31,"body":"But in the late sixties and early seventies experiments by Jeffrey Hounsfield invented a new"},{"startTime":278.31,"endTime":278.641,"body":"way of using X-rays where he, he scanned patients using a series of exposures to"},{"startTime":278.641,"endTime":278.972,"body":"a narrow beam of X-rays, which he could then reconstruct to provide cross-sectional imaging like"},{"startTime":278.972,"endTime":289.565,"body":"slices through a patient."},{"startTime":289.665,"endTime":294.986,"body":"So that's the birth of CT scanning and CT scanning has been just iteratively improved"},{"startTime":294.986,"endTime":296.405,"body":"and improved since then."},{"startTime":296.545,"endTime":302.333,"body":"And now we can scan a patient from head to toe in a matter of"},{"startTime":302.333,"endTime":306.965,"body":"moments, giving us just vast volumes of incredibly detailed and useful information."},{"startTime":307.305,"endTime":312.851,"body":"So as an example here of, of, uh, the difference in power between CT scans"},{"startTime":312.851,"endTime":313.22,"body":"and plain x-rays, we've got a, uh, a patient who, who arrived in, in the"},{"startTime":313.22,"endTime":321.725,"body":"emergency department and, and had a plain chest x-ray."},{"startTime":321.725,"endTime":325.725,"body":"They, they had, uh, chest pain and, and shortness of breath, and they were unwell."},{"startTime":326.265,"endTime":330.827,"body":"And, uh, if you look on the original image, there are some very subtle findings"},{"startTime":330.827,"endTime":333.565,"body":"about changes in vascular markings in the right lung."},{"startTime":334.105,"endTime":340.501,"body":"But actually the CT scan that they then had shows this, uh, extremely obvious sort"},{"startTime":340.501,"endTime":340.928,"body":"of, uh, um, uh, pulmonary embolus there with a big blood clot, that's a threatening"},{"startTime":340.928,"endTime":347.325,"body":"problem."},{"startTime":348.145,"endTime":353.645,"body":"So it's this precision, um, makes CT just essential to our hospital care."},{"startTime":355.705,"endTime":361.608,"body":"So other modalities have been discovered, um, like ultrasound where high frequency, uh, sound waves"},{"startTime":361.608,"endTime":366.725,"body":"are pinged in from a transducer held against the patient into the tissues."},{"startTime":366.905,"endTime":369.885,"body":"And an image is generated by listening to the reflected echoes."},{"startTime":370.425,"endTime":375.437,"body":"So we're all familiar with seeing ultrasound in the context of pregnancy scans, but ultrasounds"},{"startTime":375.437,"endTime":378.445,"body":"also a vital tool, um, for imaging cancer patients."},{"startTime":378.745,"endTime":381.085,"body":"And it's one of our main tools that we use in breast imaging."},{"startTime":382.145,"endTime":384.645,"body":"So it has the benefit of being a portable tool as well."},{"startTime":384.785,"endTime":386.965,"body":"And so that allows us to use it at the bedside."},{"startTime":387.145,"endTime":391.429,"body":"And now there are even tiny handheld units that plug into your phone, but we"},{"startTime":391.429,"endTime":394.285,"body":"use it all the time for performing image guiding interventions."},{"startTime":395.265,"endTime":399.199,"body":"So this image shows sheer wave elastography, which is one of the newer developments in"},{"startTime":399.199,"endTime":399.461,"body":"ultrasound, then that, that gives us information about the stiffness of tissue as well as"},{"startTime":399.461,"endTime":404.445,"body":"the, the shape of it."},{"startTime":404.465,"endTime":406.645,"body":"And we can also get information about blood flow."},{"startTime":407.345,"endTime":411.605,"body":"Um, and we can see that there's a small breast cancer on the right that's"},{"startTime":411.605,"endTime":414.445,"body":"really helped by the elastography to make that more obvious."},{"startTime":414.865,"endTime":419.38,"body":"And this much larger lump on, on, on the right of the screen here is"},{"startTime":419.38,"endTime":420.885,"body":"a, a benign fiber adenoma."},{"startTime":428.465,"endTime":432.997,"body":"So next, if we move on to MRI imaging, um, this was also invented in"},{"startTime":432.997,"endTime":436.925,"body":"the seventies, later on in the seventies by a physicist called Peter Mansfield."},{"startTime":437.265,"endTime":442.261,"body":"And I think this technology has a strong claim to be the most incredible technology"},{"startTime":442.261,"endTime":445.925,"body":"yet developed in any domain of human kind of endeavor, really."},{"startTime":446.265,"endTime":452.327,"body":"We, we lie a patient down inside a tube surrounded by an incredibly strong superconducting"},{"startTime":452.327,"endTime":455.965,"body":"magnet that's cooled down to minus 269 degrees C."},{"startTime":456.105,"endTime":458.085,"body":"That's four degrees above absolute zero."},{"startTime":458.985,"endTime":464.485,"body":"So this lines up the protons within the patient and we ping these protons with"},{"startTime":464.485,"endTime":464.851,"body":"a series of radio frequency pulses and then listen to the very quiet replies, and"},{"startTime":464.851,"endTime":475.485,"body":"we can generate imaging from these signals using powerful computers and mathematics to reconstruct them."},{"startTime":476.185,"endTime":479.845,"body":"And MRI also has been continuously developed since it was invented."},{"startTime":479.905,"endTime":484.263,"body":"And, and it doesn't really show any signs of stopping in terms of, uh, its"},{"startTime":484.263,"endTime":484.845,"body":"development recently."},{"startTime":485.865,"endTime":490.58,"body":"AI based image reconstruction has been deployed and that really can speed up the, uh,"},{"startTime":490.58,"endTime":490.894,"body":"the amount of time we need to keep patients in those scanners by up to"},{"startTime":490.894,"endTime":498.125,"body":"kind of four times whilst maintaining the image quality."},{"startTime":499.025,"endTime":503.541,"body":"So MRI shows us all sorts of different tissue properties depending on how the physicist"},{"startTime":503.541,"endTime":504.445,"body":"build the sequence."},{"startTime":505.425,"endTime":508.565,"body":"Um, and, and again, it generates huge volumes of data."},{"startTime":508.765,"endTime":513.044,"body":"A typical breast MRI that I might report contains almost 2000 images to review."},{"startTime":513.865,"endTime":518.051,"body":"Um, and we can see here there's an image of a, uh, a patient who"},{"startTime":518.051,"endTime":520.005,"body":"has breast cancer in their right breast."},{"startTime":521.345,"endTime":526.676,"body":"Um, and this kind of side view here, this zoomed inside view shows that it's"},{"startTime":526.676,"endTime":529.165,"body":"invading the pectoralis muscle in great detail."},{"startTime":529.225,"endTime":532.925,"body":"So that's really useful for, for the surgeons to know about before an operation."},{"startTime":533.745,"endTime":538.451,"body":"Um, and this is a, a beautiful image of, of a, a full body MRI"},{"startTime":538.451,"endTime":542.845,"body":"in coronal section as if just a slice all the way through the patient."},{"startTime":543.545,"endTime":548.292,"body":"And the on the left, we've got, sorry, the right, uh, the, the beautiful image"},{"startTime":548.292,"endTime":548.925,"body":"of Mr."},{"startTime":548.925,"endTime":551.605,"body":"Tractography, which, which shows the neural pathways in the brain."},{"startTime":551.785,"endTime":555.045,"body":"And there's loads of other ways that the MRI is use useful."},{"startTime":559.825,"endTime":564.445,"body":"So another type of imaging concentrates on the functional aspect of health and disease."},{"startTime":564.675,"endTime":569.59,"body":"This is, uh, uh, called nuclear medicine or, or, and based on radioactive nucleotides, it,"},{"startTime":569.59,"endTime":569.918,"body":"uh, sometimes rather ungenerously used to be known as unclear medicine because of the rather"},{"startTime":569.918,"endTime":570.246,"body":"sort of blobby images that you get that kind of low resolution compared to other"},{"startTime":570.246,"endTime":580.405,"body":"bits of radiology."},{"startTime":580.985,"endTime":583.045,"body":"But this, again, is changing dramatically."},{"startTime":583.045,"endTime":588.445,"body":"And there's a technology called time of flight, uh, PET imaging that's just a mind"},{"startTime":588.445,"endTime":589.165,"body":"blowing technique."},{"startTime":589.345,"endTime":595.005,"body":"So PET imaging uses radioisotopes that emit antimatter particles called positrons."},{"startTime":595.625,"endTime":598.645,"body":"Um, and that's the antimatter equivalent of electrons."},{"startTime":599.105,"endTime":602.845,"body":"So these are generated in a particle collider machine called a cyclotron."},{"startTime":603.505,"endTime":609.445,"body":"So we then bind these a positron emitters to a sugar analog mono molecule that's"},{"startTime":609.445,"endTime":613.405,"body":"designed to accumulate in highly active cells, including cancer cells."},{"startTime":614.025,"endTime":619.405,"body":"And when the positrons emitted encounter an electron, there's something called a matter anti-matter annihilation"},{"startTime":619.405,"endTime":619.763,"body":"reaction that occurs that releases the energy from those two particles, which no longer exist,"},{"startTime":619.763,"endTime":630.165,"body":"but it becomes two different gamma rays heading off in 180 degrees to each other."},{"startTime":631.305,"endTime":636.837,"body":"Um, so these gamma rays are then picked up by detectors, and it can measure"},{"startTime":636.837,"endTime":640.525,"body":"the difference in time that each one of these arrives."},{"startTime":641.185,"endTime":644.965,"body":"And then it can calculate the distance that this annihilation reaction occurred at."},{"startTime":645.145,"endTime":649.547,"body":"So these gamma rays that are moving at the speed of light and they're moving"},{"startTime":649.547,"endTime":649.841,"body":"over a distance of less than a meter, I I just find it completely incredible"},{"startTime":649.841,"endTime":656.885,"body":"that someone has created technology with that kind of precision."},{"startTime":657.345,"endTime":662.196,"body":"Um, I mean, the whole thing with pet imaging just sounds like science fiction with"},{"startTime":662.196,"endTime":666.725,"body":"anato particles and annihilation reactions, but it's happening every day throughout the developed world."},{"startTime":666.995,"endTime":667.765,"body":"It's amazing."},{"startTime":669.545,"endTime":674.003,"body":"So we can see that these, this is a, a pet scan of a, a"},{"startTime":674.003,"endTime":678.165,"body":"lady with lymphoma and it's affecting her breasts and some of her lymph nodes."},{"startTime":678.705,"endTime":683.018,"body":"Um, and we can also see some of the normal, uh, parts where the pet"},{"startTime":683.018,"endTime":687.045,"body":"tracer is accumulating in the kidneys, where it's being excreted and into the bladder."},{"startTime":687.625,"endTime":689.005,"body":"Um, and also up into the brain."},{"startTime":689.285,"endTime":691.165,"body":"'cause the brain is highly active tissue."},{"startTime":691.225,"endTime":693.565,"body":"And I hope all of yours is during the talk."},{"startTime":696.265,"endTime":699.752,"body":"So at the same time as doing the PET scan, uh, we do a CT"},{"startTime":699.752,"endTime":701.845,"body":"scan as well, and it's in, in PET ct."},{"startTime":702.105,"endTime":706.985,"body":"So we get the highly detailed information from the CT scan and the functional information"},{"startTime":706.985,"endTime":710.565,"body":"from the PET scan giving us these fantastically useful, uh, images."},{"startTime":711.545,"endTime":715.974,"body":"So we've got all these incredible sources of imaging and functional information coming from our"},{"startTime":715.974,"endTime":716.565,"body":"radiology departments."},{"startTime":716.625,"endTime":719.445,"body":"And in a way this is, uh, both a blessing and a curse."},{"startTime":719.945,"endTime":724.816,"body":"So this is, there's ever increasing pressure on radiology departments from the sheer weight of"},{"startTime":724.816,"endTime":726.765,"body":"the clinical imaging, which is required."},{"startTime":727.225,"endTime":732.245,"body":"And there's also undoubtedly some information within this imaging, which we're struggle to appreciate."},{"startTime":733.985,"endTime":737.435,"body":"So a very simple example is, is shown by the way that we have to"},{"startTime":737.435,"endTime":738.125,"body":"review CT scans."},{"startTime":738.505,"endTime":744.959,"body":"So humans are able to, um, perceive approximately 32 levels of gray, but the information"},{"startTime":744.959,"endTime":745.389,"body":"recorded with by a CT scanner includes much, much greater, uh, kind of subtlety of"},{"startTime":745.389,"endTime":753.565,"body":"a range of gray scales."},{"startTime":754.585,"endTime":759.685,"body":"Um, we therefore use what we call windowing, where we limit the range of gray"},{"startTime":759.685,"endTime":761.725,"body":"levels displayed to demonstrate different features."},{"startTime":762.145,"endTime":766.315,"body":"So for example, we might review a CT scan first on, so-called soft tissue windows,"},{"startTime":766.315,"endTime":770.485,"body":"which give us the best gray levels to show the lesions within the soft tissue."},{"startTime":770.545,"endTime":776.8,"body":"So we can see the lumps here within the liver of this, uh, cancer patient"},{"startTime":776.8,"endTime":778.885,"body":"unfortunately show, uh, liver metastasis."},{"startTime":779.705,"endTime":783.89,"body":"But if we want to see the lung lesions, we, we can't see them clearly"},{"startTime":783.89,"endTime":785.285,"body":"on this soft tissue window."},{"startTime":785.345,"endTime":790.616,"body":"So we move to lung window where we can see some lung, um, some, uh,"},{"startTime":790.616,"endTime":790.967,"body":"lung metastases at the top of the right lung there, um, and some probable bone"},{"startTime":790.967,"endTime":797.645,"body":"metastases on the bone windows."},{"startTime":798.705,"endTime":804.531,"body":"So there's a whole area of study called radios, which aims to identify and use"},{"startTime":804.531,"endTime":808.805,"body":"quantitative feature, quantitative features, uh, within imaging that might go unappreciated."},{"startTime":809.185,"endTime":812.125,"body":"So it's part of the application of AI and radiology."},{"startTime":812.225,"endTime":815.525,"body":"But I'm, I'm not gonna discuss, uh, radios in such detail."},{"startTime":815.745,"endTime":820.205,"body":"I'm gonna rather concentrate on the broader kind of image, image classification type technologies."},{"startTime":821.025,"endTime":826.157,"body":"But using computers to analyze medical images has, has a really long history, um, trying"},{"startTime":826.157,"endTime":830.605,"body":"to attempt to address some of these problems and some of these opportunities."},{"startTime":831.185,"endTime":835.819,"body":"And this has really heated up in the past few years because we've now got"},{"startTime":835.819,"endTime":836.128,"body":"so much data stored digitally in hospital systems, um, including containing information that we know"},{"startTime":836.128,"endTime":842.925,"body":"that we are not making the most of."},{"startTime":843.465,"endTime":849.575,"body":"Um, and then we've got the advancements in computing power and advancements in artificial tech,"},{"startTime":849.575,"endTime":851.205,"body":"uh, artificial intelligence techniques."},{"startTime":851.585,"endTime":855.32,"body":"So we've got really good questions to ask and we've got some really good ways"},{"startTime":855.32,"endTime":856.565,"body":"to try and answer them."},{"startTime":861.025,"endTime":865.045,"body":"So before we dive into these details, let's define a few terms that we use."},{"startTime":865.225,"endTime":869.725,"body":"So the UK National AI strategy defines AI to be machines that perform tasks normally"},{"startTime":869.725,"endTime":870.025,"body":"requiring human intelligence, especially when the machines learn from the data how to do those"},{"startTime":870.025,"endTime":874.525,"body":"tasks."},{"startTime":874.985,"endTime":878.337,"body":"And there's a guy called Stuart Russell, who's a professor at Berkeley, and he did"},{"startTime":878.337,"endTime":880.125,"body":"the BBC reflexes a couple of years ago."},{"startTime":880.205,"endTime":881.365,"body":"I I really recommend them."},{"startTime":881.825,"endTime":885.505,"body":"Um, he says that AI is about building machines that do the right thing, that"},{"startTime":885.505,"endTime":888.205,"body":"act in ways which can be expected to achieve their objectives."},{"startTime":888.625,"endTime":893.159,"body":"And then importantly, he kind of goes on to discuss how we set these objectives"},{"startTime":893.159,"endTime":896.485,"body":"and, and the possibility and dangers of miss specifying these objectives."},{"startTime":898.305,"endTime":903.405,"body":"So to define a few more terms used, um, machine learning refers to the development"},{"startTime":903.405,"endTime":903.745,"body":"of programs that are able to alter their internal parameters on the basis of, um,"},{"startTime":903.745,"endTime":912.245,"body":"of, of the data presented to them without being explicitly programmed."},{"startTime":913.025,"endTime":917.716,"body":"Neural networks are a design of computer program that's loosely based on our understanding of"},{"startTime":917.716,"endTime":918.029,"body":"how brains process information and deep learning is a type of neural network design that"},{"startTime":918.029,"endTime":924.285,"body":"refers to multiple steps or layers."},{"startTime":924.665,"endTime":929.296,"body":"And it's these type of networks that all of the, uh, recent excitement has been"},{"startTime":929.296,"endTime":929.605,"body":"about."},{"startTime":933.465,"endTime":937.08,"body":"So to look at some examples of what's been going on, I'll concentrate on my"},{"startTime":937.08,"endTime":938.285,"body":"home turf of breast imaging."},{"startTime":938.905,"endTime":942.277,"body":"Um, and let's illustrate this with an imagined patient who is a kind of composite"},{"startTime":942.277,"endTime":944.525,"body":"of many different patients I might see in my clinics."},{"startTime":945.025,"endTime":949.765,"body":"So Sheila attended for her first breast screening appointment, age 51."},{"startTime":950.385,"endTime":954.747,"body":"No con no concerns were found on her mammograms, although we did notice that she"},{"startTime":954.747,"endTime":955.038,"body":"had quite dense breasts and we know that we are more likely to miss things"},{"startTime":955.038,"endTime":960.565,"body":"with the so-called dense breasts."},{"startTime":961.545,"endTime":965.52,"body":"She diligently self examined on a regular basis, and 18 months later, she found a"},{"startTime":965.52,"endTime":966.845,"body":"lump in her right breast."},{"startTime":966.945,"endTime":970.685,"body":"So her GP referred her to the hospital breast clinic for further assessment."},{"startTime":971.585,"endTime":975.464,"body":"And there we found that she had a concerning lump on examination and we could"},{"startTime":975.464,"endTime":979.085,"body":"see that on ultrasound and on the mammogram we did that day as well."},{"startTime":979.105,"endTime":985.127,"body":"We could kind of see an area of tissue that probably, probably did, uh, correspond"},{"startTime":985.127,"endTime":985.529,"body":"to that lump, though it was difficult to appreciate 'cause it was essentially difficult to"},{"startTime":985.529,"endTime":995.165,"body":"distinguish from the surrounding tissue in that sort of situation."},{"startTime":995.165,"endTime":1000.415,"body":"We always go back to the last screening mammogram as well and scrutinize that looking"},{"startTime":1000.415,"endTime":1000.765,"body":"for, uh, whether we missed the obvious things, really, um, making sure we learn from,"},{"startTime":1000.765,"endTime":1006.365,"body":"from cases."},{"startTime":1007.065,"endTime":1011.165,"body":"Um, and, you know, possibly there was a lump there."},{"startTime":1011.905,"endTime":1016.78,"body":"It again, it looked kind of indistinguishable from the rest of the tissue and that's,"},{"startTime":1016.78,"endTime":1018.405,"body":"that's not an uncommon situation."},{"startTime":1019.305,"endTime":1022.285,"body":"So it's fair to say that best breast screening is by no means perfect."},{"startTime":1023.345,"endTime":1027.479,"body":"We only use mammograms and we only just make a binary decision about do they"},{"startTime":1027.479,"endTime":1029.685,"body":"probably have cancer at that moment or not."},{"startTime":1030.464,"endTime":1034.826,"body":"Um, so in the UK and most of Europe, we use double reading where each"},{"startTime":1034.826,"endTime":1037.444,"body":"set of mammograms is read by two different readers."},{"startTime":1037.704,"endTime":1041.69,"body":"And if there's a disagreement between them, the case is decided by a third reader"},{"startTime":1041.69,"endTime":1043.285,"body":"or, or a group of readers."},{"startTime":1044.185,"endTime":1048.338,"body":"So despite our best efforts, three outta four patients that we recall for further imaging,"},{"startTime":1048.338,"endTime":1048.615,"body":"because we are concerned about the appearance of their mammograms, they don't have cancer, they're"},{"startTime":1048.615,"endTime":1053.045,"body":"false alarms."},{"startTime":1053.985,"endTime":1059.868,"body":"And unfortunately, like Sheila, about one third of cancers that occur in screen, women aren't"},{"startTime":1059.868,"endTime":1061.045,"body":"detected by screening."},{"startTime":1061.825,"endTime":1066.743,"body":"So this room for improvement in breast screening together with the kind of high volume"},{"startTime":1066.743,"endTime":1067.071,"body":"of the imaging and the well-ordered clinical data where perhaps the objectives are a little"},{"startTime":1067.071,"endTime":1067.399,"body":"easier to specify than in many cases of radiology and medicine has this, has made"},{"startTime":1067.399,"endTime":1080.845,"body":"breast imaging a, an attractive AR area to try and apply computer analysis."},{"startTime":1083.745,"endTime":1089.421,"body":"So initial approaches used kind of basic image processing techniques and involved radiologists and computer"},{"startTime":1089.421,"endTime":1089.799,"body":"scientists working together to create mathematical descriptions of the features that we look for, um,"},{"startTime":1089.799,"endTime":1098.125,"body":"in mammograms like size and shape and density."},{"startTime":1099.145,"endTime":1104.457,"body":"So these features were then analyzed with advanced statistics and kind of classical machine learning"},{"startTime":1104.457,"endTime":1107.645,"body":"to identify these handcrafted features and to classify them."},{"startTime":1108.505,"endTime":1113.125,"body":"And this is the basis of older generation computerated detection or CAD systems."},{"startTime":1117.265,"endTime":1122.095,"body":"So these, these older CAD systems have been widely used in America, but they never"},{"startTime":1122.095,"endTime":1126.925,"body":"gained traction in Europe, perhaps because they never came close to human levels of specificity."},{"startTime":1127.275,"endTime":1129.365,"body":"That is they had too many false alarms on them."},{"startTime":1130.025,"endTime":1134.933,"body":"So in a 2015 study done in the US by Professor Lehman and her colleagues,"},{"startTime":1134.933,"endTime":1135.26,"body":"they looked at over a half a million cases and concluded that CAD does not"},{"startTime":1135.26,"endTime":1141.805,"body":"improve the diagnostic accuracy of mammography."},{"startTime":1144.625,"endTime":1150.965,"body":"So these new approaches are based on artificial in, uh, neural networks and deep learning."},{"startTime":1151.505,"endTime":1156.95,"body":"So artificial neural networks are inspired by real neurons and they have multiple inputs, some"},{"startTime":1156.95,"endTime":1157.313,"body":"kind of information processing going on in the cell body, and then an output to"},{"startTime":1157.313,"endTime":1163.485,"body":"multiple other cells."},{"startTime":1165.385,"endTime":1170.46,"body":"And a single artificial neuron has a similar configuration with lots of connections to both"},{"startTime":1170.46,"endTime":1170.798,"body":"input neurons and output neurons and some kind of very simple information processing, just a"},{"startTime":1170.798,"endTime":1177.565,"body":"sum really in the mathematical neuron."},{"startTime":1180.385,"endTime":1184.77,"body":"So these systems must be able to alter their internal parameters to enable them to"},{"startTime":1184.77,"endTime":1186.525,"body":"learn to produce a useful output."},{"startTime":1186.985,"endTime":1192.035,"body":"So each of the connection there is a weight which increases or decreases the contribution"},{"startTime":1192.035,"endTime":1193.045,"body":"of that connection."},{"startTime":1194.225,"endTime":1196.685,"body":"So we can see these are sort of roughly analogous."},{"startTime":1200.425,"endTime":1202.805,"body":"So these artificial neurons can be built into a network."},{"startTime":1203.435,"endTime":1209.285,"body":"This has an input layer, an output layer, and then hidden layers in between where"},{"startTime":1209.285,"endTime":1210.845,"body":"incremental information processing occurs."},{"startTime":1212.585,"endTime":1218.03,"body":"In our case, we might want to input a mammogram image and output a classification"},{"startTime":1218.03,"endTime":1219.845,"body":"of cancer or not cancer."},{"startTime":1220.825,"endTime":1225.76,"body":"So initially, if these weights are set randomly, the output will be non-discriminatory, then we"},{"startTime":1225.76,"endTime":1227.405,"body":"need to train the network."},{"startTime":1227.625,"endTime":1231.928,"body":"And this is done by showing the ground truth of an example to network and"},{"startTime":1231.928,"endTime":1235.085,"body":"calculating the error that the network makes compared to this truth."},{"startTime":1235.665,"endTime":1237.565,"body":"So this is known as supervised learning."},{"startTime":1237.875,"endTime":1241.485,"body":"There's a classification label that's provided to the AI system."},{"startTime":1243.945,"endTime":1250.245,"body":"So they then use very clever mathematical technique called back propagation, and that incrementally optimizes"},{"startTime":1250.245,"endTime":1254.445,"body":"the weights throughout the entire network to minimize this error."},{"startTime":1255.585,"endTime":1260.205,"body":"So the weights are adjusted to strengthen connections that leads to a useful output."},{"startTime":1260.345,"endTime":1266.633,"body":"And this is repeated and repeated and until the output is optimized, and this leads"},{"startTime":1266.633,"endTime":1271.245,"body":"to pathways and features which have stronger predictive results becoming dominant."},{"startTime":1272.105,"endTime":1275.645,"body":"And the final output of these systems will be a probability for each class."},{"startTime":1277.385,"endTime":1280.035,"body":"So we'd hope for a cancer case that we get a much higher probability that"},{"startTime":1280.035,"endTime":1280.565,"body":"it's a cancer."},{"startTime":1281.785,"endTime":1284.925,"body":"But as for human readers, there is kind of some nuance to this."},{"startTime":1285.185,"endTime":1288.968,"body":"So some cases are really easy to be definite about and some we are much"},{"startTime":1288.968,"endTime":1289.725,"body":"more uncertain of."},{"startTime":1293.385,"endTime":1298.965,"body":"So the type of network, which has been widely used recently for image classification tasks"},{"startTime":1298.965,"endTime":1302.685,"body":"is called a convolutional neural try again, convolutional neural network."},{"startTime":1303.585,"endTime":1309.753,"body":"Um, it's a design inspired also by biology, by by the biology of our visual"},{"startTime":1309.753,"endTime":1310.165,"body":"systems."},{"startTime":1310.235,"endTime":1314.205,"body":"This time having kind of receptive fields and layers of analysis."},{"startTime":1315.385,"endTime":1319.412,"body":"So the maths that's performed by the network is a convolution or a filtering, and"},{"startTime":1319.412,"endTime":1319.68,"body":"a filter is like a little piece of an image that contains a specific feature"},{"startTime":1319.68,"endTime":1324.245,"body":"or a pattern."},{"startTime":1324.585,"endTime":1326.885,"body":"In this case, I've drawn a diagonal line."},{"startTime":1327.545,"endTime":1332.375,"body":"So this filter is applied or translated across the input image representation from the previous"},{"startTime":1332.375,"endTime":1337.205,"body":"layer, and it makes an activation map or feature map of where features are identified."},{"startTime":1337.505,"endTime":1339.805,"body":"And then this map is passed onto the next layer."},{"startTime":1342.305,"endTime":1348.194,"body":"So the filters are, are described by the weights which are trained and optimized during"},{"startTime":1348.194,"endTime":1349.765,"body":"the SPAC propagation process."},{"startTime":1350.305,"endTime":1355.815,"body":"So the filters are adjusted and optimized to identify the important features that lead to"},{"startTime":1355.815,"endTime":1357.285,"body":"improved accuracy of classification."},{"startTime":1360.105,"endTime":1366.555,"body":"And as the network is trained, the initial layers might contain simple representations like lines"},{"startTime":1366.555,"endTime":1371.285,"body":"and edges and corners, but the deeper layers become increasingly abstract."},{"startTime":1372.355,"endTime":1377.777,"body":"This complexity allows the network to accurately classify the input, but it also means that"},{"startTime":1377.777,"endTime":1378.139,"body":"it becomes effectively impossible to understand kind of what's going on in these deeper layers"},{"startTime":1378.139,"endTime":1384.285,"body":"of the network."},{"startTime":1385.505,"endTime":1388.325,"body":"So various different types of layers make up these networks."},{"startTime":1388.325,"endTime":1393.845,"body":"And there's convolutional layers used for feature extraction, pooling layers to make the maths work."},{"startTime":1393.985,"endTime":1397.005,"body":"And then output layers used to perform classification."},{"startTime":1397.585,"endTime":1402.725,"body":"So the networks used for medical imaging classification might have over a hundred layers and"},{"startTime":1402.725,"endTime":1405.125,"body":"millions and millions of connections and weights."},{"startTime":1405.545,"endTime":1410.592,"body":"And it's this complexity that allows this to be a much more versatile and powerful"},{"startTime":1410.592,"endTime":1413.285,"body":"approach than the previous effort with handcrafted rules."},{"startTime":1420.425,"endTime":1425.645,"body":"So kind of that's my understanding about the fundamentals of how convolutional neural networks learn."},{"startTime":1426.555,"endTime":1428.485,"body":"Next, we need to think about how to train them."},{"startTime":1429.505,"endTime":1430.765,"body":"So they need lots of data."},{"startTime":1431.585,"endTime":1437.189,"body":"Um, publish studies and findings from employing these networks in a non-medical setting suggests that"},{"startTime":1437.189,"endTime":1439.805,"body":"basically the more data, the better, um,"},{"startTime":1441.875,"endTime":1447.445,"body":"Data sets used in the context of breast imaging, uh, typically contain tens or hundreds"},{"startTime":1447.445,"endTime":1450.045,"body":"of thousands or even millions of images."},{"startTime":1450.945,"endTime":1454.485,"body":"So the data set will be divided into training and validation and test sets."},{"startTime":1454.705,"endTime":1457.588,"body":"And the training set is used by the network to set its weights again and"},{"startTime":1457.588,"endTime":1458.165,"body":"again and again."},{"startTime":1458.905,"endTime":1464.662,"body":"The validation or tuning set is used to evaluate the performance on cases not seen"},{"startTime":1464.662,"endTime":1466.965,"body":"while training to check for generalizability."},{"startTime":1467.345,"endTime":1471.322,"body":"So the results of this will be used to monitor how well the model is"},{"startTime":1471.322,"endTime":1471.587,"body":"learning and tune the network architecture, like changing the number and order of the hidden"},{"startTime":1471.587,"endTime":1475.565,"body":"layers."},{"startTime":1478.065,"endTime":1482.325,"body":"Now, the test set shouldn't be used until you want to demonstrate the performance."},{"startTime":1483.105,"endTime":1486.845,"body":"Um, it should be totally independent from the development set ideally."},{"startTime":1487.585,"endTime":1490.445,"body":"So as an analogy, the test set is like the final exam."},{"startTime":1490.585,"endTime":1494.106,"body":"The validation set is all your past papers, which you can refer to as many"},{"startTime":1494.106,"endTime":1495.045,"body":"times as you like."},{"startTime":1495.505,"endTime":1499.836,"body":"And the training set is all your textbooks, all your notes, and all your relevant"},{"startTime":1499.836,"endTime":1500.125,"body":"experience."},{"startTime":1501.825,"endTime":1502.205,"body":"So the"},{"startTime":1502.205,"endTime":1506.514,"body":"Tests should be representative of your real world use case where you are hoping to"},{"startTime":1506.514,"endTime":1508.525,"body":"find a, find a case for this."},{"startTime":1509.305,"endTime":1514.196,"body":"So if there are biases in the validation and especially the test sets, then they"},{"startTime":1514.196,"endTime":1516.805,"body":"may, these systems may not perform as expected."},{"startTime":1518.425,"endTime":1521.885,"body":"So training data is vital and we'll come back to that right at the end."},{"startTime":1522.385,"endTime":1526.657,"body":"So there must be lots of it, it must be reliable and the outcomes that"},{"startTime":1526.657,"endTime":1530.645,"body":"you're hoping the model to help with must be accurately represented in this data."},{"startTime":1536.625,"endTime":1538.805,"body":"So kind of where are we up to now?"},{"startTime":1539.465,"endTime":1542.535,"body":"I'd, I'd like to just have a think about how we look at performance of"},{"startTime":1542.535,"endTime":1545.605,"body":"these systems and where we're up to in terms of the research and the use."},{"startTime":1546.335,"endTime":1548.245,"body":"Again, I'm gonna concentrate on breast screening."},{"startTime":1549.225,"endTime":1554.009,"body":"So this paper from 2020 was the first to do a really good comparison of"},{"startTime":1554.009,"endTime":1555.285,"body":"breast screening AI systems."},{"startTime":1556.345,"endTime":1560.765,"body":"You can see that they present the performance of these systems as curves."},{"startTime":1560.765,"endTime":1563.165,"body":"These are known as receiver operating characteristic curves."},{"startTime":1563.675,"endTime":1568.672,"body":"Essentially these plot how good the system is at correctly classifying cancers against how many"},{"startTime":1568.672,"endTime":1570.005,"body":"false alarms it causes."},{"startTime":1570.465,"endTime":1577.112,"body":"And that difference occurs as you change what probability threshold you use for calling a"},{"startTime":1577.112,"endTime":1578.885,"body":"case a probable cancer."},{"startTime":1579.505,"endTime":1582.405,"body":"So the ideal performance will be going straight up to the top left corner."},{"startTime":1582.405,"endTime":1583.325,"body":"And then across the top,"},{"startTime":1584.855,"endTime":1590.128,"body":"These were, these curves work really well from comparing ang algorithms as they're under development"},{"startTime":1590.128,"endTime":1590.479,"body":"or is in the case of this study to compare different algorithms providing you use"},{"startTime":1590.479,"endTime":1590.831,"body":"it on precisely the same set of cases, but if you want to compare to,"},{"startTime":1590.831,"endTime":1605.245,"body":"uh, human readers with human readers, you won't, you can't get a curve."},{"startTime":1605.945,"endTime":1609.125,"body":"Um, you'll only get one value per reader, really."},{"startTime":1609.305,"endTime":1611.165,"body":"And that gained from looking at lots of cases."},{"startTime":1611.625,"endTime":1615.753,"body":"And actually this is usually averaged over lots of readers in order to reduce the"},{"startTime":1615.753,"endTime":1617.405,"body":"margins of error in the statistics."},{"startTime":1618.785,"endTime":1624.095,"body":"Um, so with this study, if we add the values of sensitivity, um, and specificity"},{"startTime":1624.095,"endTime":1624.449,"body":"from the average human performance for the first reader and the second reader, and then"},{"startTime":1624.449,"endTime":1624.803,"body":"this consensus thing that, that's what I was talking about with the, with the double"},{"startTime":1624.803,"endTime":1625.157,"body":"reading in, in, uh, breast screening, you find that the, the performance is kind of,"},{"startTime":1625.157,"endTime":1625.511,"body":"kind of close to the first and second reader, but can't quite, uh, reach the"},{"startTime":1625.511,"endTime":1646.045,"body":"consensus opinion."},{"startTime":1647.625,"endTime":1651.085,"body":"And this study that I was involved with found had very similar findings."},{"startTime":1651.085,"endTime":1656.317,"body":"This was also in 2020 where uh, it looked like it probably outperformed the first"},{"startTime":1656.317,"endTime":1659.805,"body":"reader similar to the second, but couldn't quite read consensus."},{"startTime":1660.145,"endTime":1664.813,"body":"And we have to remember that consensus, uh, the consensus opinion is our standard of"},{"startTime":1664.813,"endTime":1665.125,"body":"care."},{"startTime":1665.985,"endTime":1670.62,"body":"So, so it looks like probably the performance is inching up a little bit in"},{"startTime":1670.62,"endTime":1672.165,"body":"the, in in the interim."},{"startTime":1672.185,"endTime":1678.005,"body":"But basically human, human expert, human and machine performance are similar."},{"startTime":1678.115,"endTime":1683.446,"body":"Perhaps on average machine performance is now beginning to, to, uh, outperform a single expert,"},{"startTime":1683.446,"endTime":1686.645,"body":"but it's something in the same kind of range."},{"startTime":1687.745,"endTime":1692.917,"body":"Now these studies have been using retrospective data and this approach has several strengths including"},{"startTime":1692.917,"endTime":1696.365,"body":"access to really large volumes of data from hospital systems."},{"startTime":1696.825,"endTime":1701.573,"body":"And that contains real clinical results of what happened, both in terms of the radiologist"},{"startTime":1701.573,"endTime":1706.005,"body":"interpretation at the time, um, and then crucially in terms of longer term outcomes."},{"startTime":1706.305,"endTime":1711.429,"body":"So both for interval cancers that present symptomatically like our imaginary patient Sheila did, and"},{"startTime":1711.429,"endTime":1714.845,"body":"also for cancers detected at the next round of screening."},{"startTime":1718.065,"endTime":1720.885,"body":"So there are also a few reasons to be cautious about this approach."},{"startTime":1721.155,"endTime":1726.271,"body":"It's inherently a look at older data and therefore it may not truly represent the"},{"startTime":1726.271,"endTime":1730.365,"body":"performance of a system on current screening images that are required today."},{"startTime":1730.665,"endTime":1733.965,"body":"And obviously that's what we actually need to use in practice."},{"startTime":1734.865,"endTime":1740.371,"body":"Um, and it doesn't capture the kind of result of the human interaction with that"},{"startTime":1740.371,"endTime":1740.738,"body":"machine that will be vital to check that we can correctly act on these uh,"},{"startTime":1740.738,"endTime":1746.245,"body":"systems."},{"startTime":1748.345,"endTime":1755.605,"body":"But happily, there is now increasing positive evidence from, uh, real life prospective work."},{"startTime":1756.025,"endTime":1762.12,"body":"Um, and, and there's been three papers out of of Scandinavia just over the last"},{"startTime":1762.12,"endTime":1764.965,"body":"few months, um, showing really promising work."},{"startTime":1765.345,"endTime":1770.853,"body":"So this, this paper published over the summer showing their preliminary results on the MAI"},{"startTime":1770.853,"endTime":1774.525,"body":"trial, and that's been done at four sites in Sweden."},{"startTime":1775.475,"endTime":1781.092,"body":"They've looked at modifying the normal double read process whereby cancer, um, ca cases are"},{"startTime":1781.092,"endTime":1782.965,"body":"considered by the AI system."},{"startTime":1783.745,"endTime":1786.245,"body":"Um, and, and a risk is attached to them."},{"startTime":1786.905,"endTime":1790.896,"body":"Now, if they're in the lowest 90% of, of the risk according to the AI"},{"startTime":1790.896,"endTime":1791.162,"body":"system, they have only a single human read, but the human read does have access"},{"startTime":1791.162,"endTime":1796.485,"body":"to the AI output as well."},{"startTime":1796.985,"endTime":1800.805,"body":"And if they're in the highest 10%, they have a double human read."},{"startTime":1801.265,"endTime":1804.845,"body":"Um, and also both readers have access to the AI output."},{"startTime":1805.705,"endTime":1810.51,"body":"So these pre preliminary results have shown a small increase in the number of cancers"},{"startTime":1810.51,"endTime":1810.831,"body":"detected with no increase in the number of false alarms, and overall a 44% reduction"},{"startTime":1810.831,"endTime":1811.151,"body":"in the number of, uh, mammograms that need to be read because of this change"},{"startTime":1811.151,"endTime":1822.045,"body":"from double reading to single reading."},{"startTime":1822.785,"endTime":1828.419,"body":"And it's also, I think, really important and interesting that it's very well accepted by"},{"startTime":1828.419,"endTime":1828.795,"body":"the screening population with only a, a fraction of 1% declining, uh, to participate in"},{"startTime":1828.795,"endTime":1834.805,"body":"this study."},{"startTime":1837.745,"endTime":1842.084,"body":"So in terms of what's going on in wider radiology, well, there are loads of"},{"startTime":1842.084,"endTime":1842.373,"body":"informa interesting research papers coming out all the time as showing all kinds of novel"},{"startTime":1842.373,"endTime":1842.662,"body":"findings, like being able to accurately predict valvular disease from chest x-rays in a way"},{"startTime":1842.662,"endTime":1842.952,"body":"that we wouldn't, we in some ways we, we know that we can do that,"},{"startTime":1842.952,"endTime":1843.241,"body":"but, uh, certainly looking more impressive than what we can achieve just by looking at"},{"startTime":1843.241,"endTime":1862.045,"body":"them in terms of products available for clinical use."},{"startTime":1862.045,"endTime":1865.245,"body":"There's this website from Radboud University that's really helpful."},{"startTime":1865.785,"endTime":1869.965,"body":"It lists all the AI type interpretation products which have been awarded a CE marking."},{"startTime":1870.265,"endTime":1874.424,"body":"So last time I checked, uh, there were 220 products on this website and there's"},{"startTime":1874.424,"endTime":1876.365,"body":"a similar one, uh, for the us."},{"startTime":1877.425,"endTime":1880.765,"body":"Um, now this covers most areas of radiology."},{"startTime":1881.025,"endTime":1885.819,"body":"So, so far most of these products seem to be quite narrow in their scope"},{"startTime":1885.819,"endTime":1886.139,"body":"and many aim to kind of give us improved efficiencies and tasks that we can"},{"startTime":1886.139,"endTime":1886.458,"body":"already do, like, uh, monitoring lesions over time, but quite a few offer capabilities which"},{"startTime":1886.458,"endTime":1898.925,"body":"we're not able to do explicitly ourselves, like quantitative risk prediction."},{"startTime":1906.265,"endTime":1912.418,"body":"So onto the next bit of my talk, as you might be aware, there's now"},{"startTime":1912.418,"endTime":1912.828,"body":"the next generation of AI developments arriving and it might potentially render most of what"},{"startTime":1912.828,"endTime":1923.085,"body":"I've just discussed, uh, obsolete or, or certainly, uh, change it."},{"startTime":1924.025,"endTime":1928.241,"body":"So this is the advent of the so-called large language models that, that are based"},{"startTime":1928.241,"endTime":1929.085,"body":"on transform networks."},{"startTime":1929.965,"endTime":1934.734,"body":"Hollywood has kindly named, uh, one of their latest films, transformers Rise of The Beasts,"},{"startTime":1934.734,"endTime":1935.052,"body":"um, and even written it on the side of a bus, which, uh, gives us"},{"startTime":1935.052,"endTime":1942.365,"body":"a nice slide to start off this section with."},{"startTime":1942.365,"endTime":1946.786,"body":"And perhaps a question for researchers like me is, is, is do we try and"},{"startTime":1946.786,"endTime":1947.965,"body":"get on that bus?"},{"startTime":1948.705,"endTime":1953.457,"body":"Um, it looks like they're gonna be central for developments in imaging AI and, and"},{"startTime":1953.457,"endTime":1954.725,"body":"healthcare AI in general."},{"startTime":1955.625,"endTime":1959.685,"body":"So again, I'm gonna try and look at how this approach functions under the bonnet."},{"startTime":1961.585,"endTime":1967.177,"body":"So transformer models were proposed by Google researchers in 2017 as an efficient and powerful"},{"startTime":1967.177,"endTime":1967.55,"body":"method, originally demonstrated for translation from one language to another, um, that they've become the"},{"startTime":1967.55,"endTime":1967.922,"body":"basis of these large language models, which as far as I can tell, have surprised"},{"startTime":1967.922,"endTime":1968.295,"body":"everyone including their designers for the kind of power, power and capabilities that see these"},{"startTime":1968.295,"endTime":1985.445,"body":"systems seem to have."},{"startTime":1986.835,"endTime":1991.757,"body":"So they're based on the idea of self supervised learning originally from text, but now"},{"startTime":1991.757,"endTime":1994.055,"body":"also using multimodal data such as images."},{"startTime":1994.835,"endTime":1998.908,"body":"So part of the power of this approach is that they can ingest a piece"},{"startTime":1998.908,"endTime":2001.895,"body":"of information and use the trick of guess what comes next."},{"startTime":2003.355,"endTime":2005.335,"body":"Um, and that can inform their learning."},{"startTime":2005.555,"endTime":2009.838,"body":"So all they need to do is to mask out what word or piece of"},{"startTime":2009.838,"endTime":2010.695,"body":"information comes next."},{"startTime":2011.285,"endTime":2015.055,"body":"Make a guess, compare that guess to what's really there when you unmask it."},{"startTime":2015.315,"endTime":2018.386,"body":"And then update the weights to minimize this error in a similar way to, we"},{"startTime":2018.386,"endTime":2019.615,"body":"kind of saw with the other."},{"startTime":2020.525,"endTime":2023.695,"body":"This means that the training sets are, are kind of everything."},{"startTime":2023.695,"endTime":2026.917,"body":"They're enormous, all of the text on the information, uh, all of the text on"},{"startTime":2026.917,"endTime":2029.495,"body":"the internet and all the books they can get their hands on."},{"startTime":2029.755,"endTime":2032.295,"body":"Um, that's what they've used this."},{"startTime":2032.345,"endTime":2036.215,"body":"Guess what next needs to be based on kind of context."},{"startTime":2036.955,"endTime":2041.659,"body":"So these models have to learn and refine their representations of language and concepts in"},{"startTime":2041.659,"endTime":2043.855,"body":"order to be able to do that."},{"startTime":2043.855,"endTime":2045.575,"body":"Guess what comes next accurately."},{"startTime":2049.275,"endTime":2053.555,"body":"So let's use this example a sentence, uh, which is a slightly strange sentence, but"},{"startTime":2053.555,"endTime":2056.695,"body":"it allows me to show off some of my children's artworks."},{"startTime":2056.695,"endTime":2056.975,"body":"Do you"},{"startTime":2057.415,"endTime":2062.415,"body":"\u003claugh\u003e?"},{"startTime":2060.235,"endTime":2065.679,"body":"So the system will align, will, the system will assign a numerical value to the"},{"startTime":2065.679,"endTime":2067.495,"body":"words it's known as tokenization."},{"startTime":2067.835,"endTime":2070.775,"body":"And that's using some kind of learned dictionary approach."},{"startTime":2071.395,"endTime":2075.007,"body":"And it will also record the pos the position of the words in the sequence"},{"startTime":2075.007,"endTime":2077.415,"body":"in a much more complicated way than I've written here."},{"startTime":2078.995,"endTime":2083.018,"body":"It will then generate a series of numbers called a vector, which can be thought"},{"startTime":2083.018,"endTime":2083.286,"body":"of as a coordinate or an address for how this word should be represented in"},{"startTime":2083.286,"endTime":2089.455,"body":"some kind of meaning space known as an embedding."},{"startTime":2090.114,"endTime":2095.215,"body":"So it calculates this depending on an attention mechanism that it learns."},{"startTime":2095.315,"endTime":2100.91,"body":"So it learns the significant associations within sentences or longer pieces of text to give"},{"startTime":2100.91,"endTime":2102.775,"body":"the words themselves proper context."},{"startTime":2103.995,"endTime":2107.775,"body":"So the model architecture here is shown on the left of the screen."},{"startTime":2108.315,"endTime":2113.855,"body":"Um, and as you can see, even though it's a simplified diagram, it's quite complex."},{"startTime":2116.825,"endTime":2122.055,"body":"Again, in a similar way to the multi-layered approach, uh, used by convolutional neural networks."},{"startTime":2122.425,"endTime":2124.495,"body":"These transformers have multiple layers."},{"startTime":2125.395,"endTime":2130.178,"body":"And again, this leads to an increasingly abstract and complex representation of this data that"},{"startTime":2130.178,"endTime":2131.135,"body":"they're presented with."},{"startTime":2131.555,"endTime":2135.014,"body":"So there's also a decode, a part of this network which extracts this information using"},{"startTime":2135.014,"endTime":2135.244,"body":"some of the information, using some of the things it's learned to encode it in"},{"startTime":2135.244,"endTime":2139.165,"body":"the first place."},{"startTime":2139.665,"endTime":2144.817,"body":"And then the output is used as the feedback to alter the weights and improve"},{"startTime":2144.817,"endTime":2147.565,"body":"the accuracy in predicting guess what comes next."},{"startTime":2150.825,"endTime":2154.517,"body":"So if we sem, if we represent this meaning space as a simple 2D graph,"},{"startTime":2154.517,"endTime":2154.763,"body":"we can show that the cat has been placed close to the dog in the"},{"startTime":2154.763,"endTime":2155.01,"body":"X value and that the cat is close to the milk in the Y direction,"},{"startTime":2155.01,"endTime":2164.365,"body":"but the, the milk and the dog are further apart."},{"startTime":2164.945,"endTime":2170.907,"body":"So if we then add a third dimension, then all of these words might be"},{"startTime":2170.907,"endTime":2174.485,"body":"related as words that children first learn to read."},{"startTime":2174.705,"endTime":2177.125,"body":"So they have similar values in the Z direction."},{"startTime":2178.425,"endTime":2183.206,"body":"So I can't carry on this similarly any further 'cause I can't visualize after three"},{"startTime":2183.206,"endTime":2183.525,"body":"dimensions."},{"startTime":2183.945,"endTime":2190.02,"body":"But if we simply think of a list of coordinates, then you can imagine this"},{"startTime":2190.02,"endTime":2192.045,"body":"extending into multiple dimensions mathematically."},{"startTime":2192.985,"endTime":2199.452,"body":"Um, but whilst maintaining a kind of meaningful relationship between the information coded for in"},{"startTime":2199.452,"endTime":2199.883,"body":"these coordinates, so apparently chat GPT, sorry GPT-3, which is now an obsolete model, has"},{"startTime":2199.883,"endTime":2211.525,"body":"about, has just over 12,000 dimensions in this, uh, in this embedding space."},{"startTime":2215.425,"endTime":2219.872,"body":"You know, another dimension with more abstract concepts might include where pet cats are related"},{"startTime":2219.872,"endTime":2222.245,"body":"to factors leading to the loss of wildlife."},{"startTime":2223.425,"endTime":2229.141,"body":"Anyway, these, so-called foundation models can then be fine tuned for specific tasks with smaller"},{"startTime":2229.141,"endTime":2230.285,"body":"targeted data sets."},{"startTime":2230.985,"endTime":2235.83,"body":"Um, so perhaps in the medical domain and a variety of, uh, methods are then"},{"startTime":2235.83,"endTime":2236.153,"body":"used to make them behave in a specific way, such as is done with reinforcement"},{"startTime":2236.153,"endTime":2236.476,"body":"learning from human feedback that they use to make a conversational chat bot like chat"},{"startTime":2236.476,"endTime":2245.845,"body":"GPT."},{"startTime":2248.105,"endTime":2252.485,"body":"So already these systems are at kind of genius level for language tasks."},{"startTime":2253.435,"endTime":2258.342,"body":"They've got a massive working memory and they show what is known as emergent skills"},{"startTime":2258.342,"endTime":2258.669,"body":"where they can handle tasks that they've never encountered in training and they can now"},{"startTime":2258.669,"endTime":2264.885,"body":"handle multimodal data as well."},{"startTime":2265.385,"endTime":2269.468,"body":"So overall, does that sound a little bit like what we hope to do when"},{"startTime":2269.468,"endTime":2270.285,"body":"we practice medicine?"},{"startTime":2271.165,"endTime":2276.634,"body":"Possibly on the downside, we also know that they may be biased, they might make"},{"startTime":2276.634,"endTime":2280.645,"body":"stuff up that's blatantly untrue, um, and have major reasoning errors."},{"startTime":2281.265,"endTime":2285.915,"body":"Um, and at the same time they'll sound scarily credible because they're grammar and language"},{"startTime":2285.915,"endTime":2286.845,"body":"will be perfect."},{"startTime":2287.785,"endTime":2291.967,"body":"And if we're feeling introspective, it's also sounds a bit familiar to what we do"},{"startTime":2291.967,"endTime":2292.525,"body":"as medics."},{"startTime":2293.025,"endTime":2297.125,"body":"Um, but it's clear whatever, that these are really powerful technologies."},{"startTime":2297.745,"endTime":2301.485,"body":"Um, and some people are even claiming it's the start of artificial general intelligence."},{"startTime":2302.385,"endTime":2306.488,"body":"So the way this information is handled by these multiple layers of abstraction means that"},{"startTime":2306.488,"endTime":2310.045,"body":"it's so complex that, and nobody understands really what's going on in there."},{"startTime":2311.145,"endTime":2314.845,"body":"But I think it's fascinating how these work and what the implications for these are."},{"startTime":2316.025,"endTime":2320.385,"body":"Um, forgive me for going off into a bit of blatant speculation for the next"},{"startTime":2320.385,"endTime":2320.676,"body":"bit, but at some level they must be generating some kind of reasonably accurate model"},{"startTime":2320.676,"endTime":2328.525,"body":"of the world just from the information that they've scraped from the internet."},{"startTime":2329.145,"endTime":2333.358,"body":"And if this is possible, then presumably they can also be persuaded to make realistic"},{"startTime":2333.358,"endTime":2335.325,"body":"models of humans in health and disease."},{"startTime":2336.065,"endTime":2339.125,"body":"So the possibilities are pretty exciting, I think."},{"startTime":2340.225,"endTime":2344.925,"body":"So people have talked about the idea of precision medicine for a long time, and"},{"startTime":2344.925,"endTime":2345.238,"body":"whilst we're inching towards that, these type of technologies could really deliver a transformational change"},{"startTime":2345.238,"endTime":2350.565,"body":"towards achieving this."},{"startTime":2350.645,"endTime":2355.672,"body":"I think, uh, this diagram is from a paper published by my colleagues at the"},{"startTime":2355.672,"endTime":2360.365,"body":"Royal Marsden regarding our ambition to move towards a integrated diagnostics and discovery system."},{"startTime":2362.075,"endTime":2367.025,"body":"This illustrates the strands of data about a patient like pathology and radiology and genomics"},{"startTime":2367.025,"endTime":2371.645,"body":"and medical records that are currently not exploited as we'd hoped they could be."},{"startTime":2371.865,"endTime":2376.596,"body":"And the idea that these could be woven together to deliver both improved care for"},{"startTime":2376.596,"endTime":2378.805,"body":"the patient and also new research insights."},{"startTime":2379.705,"endTime":2383.662,"body":"So this path per is about the kind of the data and the idea and"},{"startTime":2383.662,"endTime":2385.245,"body":"the possibilities of joining this paper."},{"startTime":2385.395,"endTime":2390.896,"body":"This data together is not about these new AI systems, but I think, you know,"},{"startTime":2390.896,"endTime":2391.263,"body":"it seems to me like these could go together very nicely being sort of rather"},{"startTime":2391.263,"endTime":2396.765,"body":"speculative."},{"startTime":2396.765,"endTime":2402.325,"body":"Again, this, this type of technology could allow us to understand the messy and complex"},{"startTime":2402.325,"endTime":2402.696,"body":"information that's routinely collected in clinical care and approach questions that we currently aren't able"},{"startTime":2402.696,"endTime":2411.965,"body":"to answer because of this complexity and even make new discoveries."},{"startTime":2412.145,"endTime":2418.597,"body":"So essentially kind of enrolling everyone in an observational clinical trial, you could envisage such"},{"startTime":2418.597,"endTime":2419.027,"body":"a technology having implications throughout the PA pathway, for example, suggesting personalized screening based on"},{"startTime":2419.027,"endTime":2419.458,"body":"family history, general healthcare information, and possibly genomics in addition to using the imaging in"},{"startTime":2419.458,"endTime":2435.805,"body":"a more advanced way than just cancer or no cancer."},{"startTime":2436.345,"endTime":2441.445,"body":"So if we go back to Sheila, you know, more perhaps a system that could"},{"startTime":2441.445,"endTime":2441.785,"body":"have suggested she needed additional imaging using a different modality might have detected her cancer"},{"startTime":2441.785,"endTime":2446.885,"body":"earlier."},{"startTime":2448.185,"endTime":2453.597,"body":"So once a cancer has been diagnosed, we might offer kind of accurate tailored prognostic"},{"startTime":2453.597,"endTime":2456.845,"body":"information and guidance on what treatments might be best."},{"startTime":2456.985,"endTime":2459.405,"body":"So going back to that idea of death, what comes next?"},{"startTime":2459.755,"endTime":2463.805,"body":"It's exactly what we're trying to do when we're making these decisions about patient's treatment."},{"startTime":2465.385,"endTime":2470.84,"body":"So for example, a model that can combine imaging from both radiology and pathology might"},{"startTime":2470.84,"endTime":2475.205,"body":"be able to more accurately predict the utility of different chemotherapy regimes."},{"startTime":2476.025,"endTime":2480.975,"body":"Um, if we add in information from the electronic health record, we could also predict"},{"startTime":2480.975,"endTime":2481.305,"body":"the toxicity risks for that patient and that might help us tailor further the, the"},{"startTime":2481.305,"endTime":2487.245,"body":"kind of treatment recommendations."},{"startTime":2488.425,"endTime":2494.19,"body":"So it could be useful for helping us with surgical planning or, or making, uh,"},{"startTime":2494.19,"endTime":2494.575,"body":"really effective and efficient post-treatment surveillance regimes rather than just sticking to a, to a"},{"startTime":2494.575,"endTime":2500.725,"body":"standard model."},{"startTime":2501.585,"endTime":2507.11,"body":"So ultimately everyone attending hospital could benefit from this kind of individualized precision approach and"},{"startTime":2507.11,"endTime":2507.479,"body":"in turn their data could be used to update models and improve performance for future"},{"startTime":2507.479,"endTime":2513.005,"body":"patients."},{"startTime":2516.505,"endTime":2520.532,"body":"So before I finish and we move on to a kind of question and answer"},{"startTime":2520.532,"endTime":2520.8,"body":"session, I wanna get some questions and concerns in there and perhaps we can all"},{"startTime":2520.8,"endTime":2525.365,"body":"discuss these together."},{"startTime":2526.505,"endTime":2526.725,"body":"Um,"},{"startTime":2529.515,"endTime":2533.817,"body":"Reliability of these systems needs to be demonstrated and we can do that both with"},{"startTime":2533.817,"endTime":2536.685,"body":"retrospective data and with prospective studies as we've discussed already."},{"startTime":2537.145,"endTime":2542.367,"body":"But then we must continuously monitor performance of these systems and look actively look for"},{"startTime":2542.367,"endTime":2544.805,"body":"adverse outcomes or drift in performance occurring."},{"startTime":2545.305,"endTime":2550.117,"body":"For example, we know that changes to a mammography machine will produce differences to an"},{"startTime":2550.117,"endTime":2553.325,"body":"AI performance if it's been trained on the older version."},{"startTime":2554.665,"endTime":2559.725,"body":"Um, it might require recalibration or it might require additional training to overcome this."},{"startTime":2560.625,"endTime":2565.489,"body":"And this raises an interesting question for, you know, when a new mammography system or"},{"startTime":2565.489,"endTime":2565.814,"body":"or any system that's been a similar context, if you see what I mean, when"},{"startTime":2565.814,"endTime":2566.138,"body":"we've got a new system introduced, we'll have no prior prior cases to learn from"},{"startTime":2566.138,"endTime":2577.165,"body":"or or to test performance on."},{"startTime":2578.345,"endTime":2583.788,"body":"So yeah, it could put us into quite a difficult position and particularly in the"},{"startTime":2583.788,"endTime":2584.151,"body":"con context of our other unresourced NHSI can envisage a situation where we no longer"},{"startTime":2584.151,"endTime":2584.514,"body":"have sufficient trained people to return to a doubled reading, a human double read without"},{"startTime":2584.514,"endTime":2595.765,"body":"any AI assistance."},{"startTime":2596.385,"endTime":2599.405,"body":"Um, and so do we remain stuck with the old system?"},{"startTime":2599.745,"endTime":2603.805,"body":"Do we have to tolerate periods of increased uncertainty about performance?"},{"startTime":2604.545,"endTime":2609.234,"body":"Should we just keep our fingers crossed that further research and development will give us"},{"startTime":2609.234,"endTime":2610.485,"body":"robustness of AI systems?"},{"startTime":2610.635,"endTime":2614.427,"body":"That means that we don't have to worry about this problem so much or should"},{"startTime":2614.427,"endTime":2614.68,"body":"we wait until that's actually been properly developed and delivered before we start to deploy"},{"startTime":2614.68,"endTime":2618.725,"body":"these systems?"},{"startTime":2620.105,"endTime":2625.567,"body":"And overreliance on these systems in general could make our hospital systems much more prone"},{"startTime":2625.567,"endTime":2628.845,"body":"to causing serious harm when it, it failures occur."},{"startTime":2629.225,"endTime":2634.005,"body":"And we'd hope that as we get more advanced, it failures become less common."},{"startTime":2634.345,"endTime":2638.74,"body":"But I don't know, in my experience it seems like they're becoming more common for"},{"startTime":2638.74,"endTime":2641.085,"body":"all sorts of just crazy, quite trivial reasons."},{"startTime":2641.185,"endTime":2646.957,"body":"We can just knock hospital systems out for half a day or a day or,"},{"startTime":2646.957,"endTime":2650.805,"body":"uh, yeah, uh, so that it's a real danger there."},{"startTime":2652.705,"endTime":2652.925,"body":"Um,"},{"startTime":2655.145,"endTime":2658.125,"body":"The human machine interaction is kind of key to this."},{"startTime":2658.435,"endTime":2663.841,"body":"There's an inherent black box element to all these systems where it's usually not possible"},{"startTime":2663.841,"endTime":2667.445,"body":"to understand explicitly how these machines have made their decisions."},{"startTime":2667.875,"endTime":2671.852,"body":"There's lots of research into this and I'm not so worried about this aspect because"},{"startTime":2671.852,"endTime":2675.565,"body":"I think partly by familiarity with these systems, we'll get better at using them."},{"startTime":2676.265,"endTime":2680.42,"body":"Um, and partly because I think there will be advancement in explainable AI because there"},{"startTime":2680.42,"endTime":2681.805,"body":"is lots of effort there."},{"startTime":2682.465,"endTime":2685.165,"body":"And after all, I don't really understand what's going on in my head."},{"startTime":2685.165,"endTime":2688.365,"body":"Well, I don't understand at all what's going on in my head in terms of"},{"startTime":2688.365,"endTime":2690.925,"body":"how I make the judgment that I'm worried enough about that patient."},{"startTime":2691.385,"endTime":2696.019,"body":"But I do learn how to explain what I've seen on those scans to my"},{"startTime":2696.019,"endTime":2696.328,"body":"colleagues so I can convince them that, uh, of, of what I've seen and they"},{"startTime":2696.328,"endTime":2703.125,"body":"can decide whether or not I'm spouting nonsense."},{"startTime":2703.985,"endTime":2709.437,"body":"Um, and at the moment we're, we're at this area, I think a relatively comfortable"},{"startTime":2709.437,"endTime":2709.801,"body":"position where I think in almost all circumstances the best way to achieve best performance"},{"startTime":2709.801,"endTime":2718.525,"body":"is to have a human and AI system working together."},{"startTime":2719.665,"endTime":2724.608,"body":"And this is also what patients are happiest with from, from what I've seen from"},{"startTime":2724.608,"endTime":2727.245,"body":"some, uh, from some, um, patient involvement events."},{"startTime":2727.915,"endTime":2733.078,"body":"However, I don't see any reason why the performance of these systems won't continue to"},{"startTime":2733.078,"endTime":2733.422,"body":"improve further and that puts us into the uncomfortable position that as if, if I"},{"startTime":2733.422,"endTime":2733.766,"body":"as a clinician intervene or overrule the AI decision, then on average I'll just introduce"},{"startTime":2733.766,"endTime":2734.111,"body":"error and I'll make outcomes worse for my patients overall, though of course it might"},{"startTime":2734.111,"endTime":2734.455,"body":"be that for that particular patient on that particular day I am right and the"},{"startTime":2734.455,"endTime":2754.765,"body":"AI is not."},{"startTime":2756.145,"endTime":2762.925,"body":"Um, I think there's a issue about kind of in some situations too much information."},{"startTime":2764.305,"endTime":2767.285,"body":"Uh, there's ignorance, it's bliss, isn't there."},{"startTime":2767.345,"endTime":2771.764,"body":"One, one likely outcome from all this is we'll be able to give patients lots"},{"startTime":2771.764,"endTime":2775.005,"body":"more and lots more accurate information about their risks and prognosis."},{"startTime":2775.785,"endTime":2778.445,"body":"Um, and this is not something that we're all that familiar with, really."},{"startTime":2778.905,"endTime":2782.685,"body":"We just dunno how people will respond to this, both clinicians and patients."},{"startTime":2782.985,"endTime":2789.165,"body":"So currently in clinical genetics, there are very careful conversations that happen pre-testing to forewarn"},{"startTime":2789.165,"endTime":2793.285,"body":"patients that their tests might show results with profound consequences."},{"startTime":2793.825,"endTime":2796.445,"body":"And that leads some people to choose not to have testing at all."},{"startTime":2796.825,"endTime":2798.805,"body":"And in some situations that's probably a good choice."},{"startTime":2798.965,"endTime":2801.405,"body":"I think that's just something we need to be aware of."},{"startTime":2803.265,"endTime":2808.027,"body":"And something that I'm a bit obsessed with is kind of, uh, is this aspect"},{"startTime":2808.027,"endTime":2810.885,"body":"of who owns what and control of these systems."},{"startTime":2811.545,"endTime":2816.723,"body":"And as we become, as all become more pronounced in, in other areas of society"},{"startTime":2816.723,"endTime":2819.485,"body":"where AI impacts our lives, I'm not ideal."},{"startTime":2819.825,"endTime":2824.861,"body":"I'm not wild about the idea that adoption of these systems could cause a centralization"},{"startTime":2824.861,"endTime":2826.205,"body":"of power and control."},{"startTime":2827.385,"endTime":2832.485,"body":"Um, and this funding and control is such a interesting subject for healthcare in my"},{"startTime":2832.485,"endTime":2833.845,"body":"opinion at the moment."},{"startTime":2834.145,"endTime":2839.118,"body":"Uh, research organizations and AI systems tend to be granted access to anonymized patient data"},{"startTime":2839.118,"endTime":2842.765,"body":"under a var variety of agreements, usually for time limited periods."},{"startTime":2843.145,"endTime":2846.182,"body":"And these, they develop products from them and these are owned by the companies who"},{"startTime":2846.182,"endTime":2848.005,"body":"will then need to make some money from them."},{"startTime":2849.025,"endTime":2852.845,"body":"Uh, in some ways these systems are a distillation of the data presented to them."},{"startTime":2853.035,"endTime":2858.645,"body":"They're a very complex digital filter to make these products work as well as possible."},{"startTime":2858.945,"endTime":2862.005,"body":"For all of us, as much data as possible needs to be available."},{"startTime":2862.745,"endTime":2867.903,"body":"And if essentially the whole population is contributing to this work, we should, should, I"},{"startTime":2867.903,"endTime":2868.246,"body":"dunno, should we all somehow have a say in and a control in in what's"},{"startTime":2868.246,"endTime":2873.405,"body":"produced?"},{"startTime":2874.145,"endTime":2874.645,"body":"I'm not sure."},{"startTime":2876.545,"endTime":2882.913,"body":"So to sum up, um, while there are huge concerns and problems to be carefully"},{"startTime":2882.913,"endTime":2883.338,"body":"addressed, overall I'm, I'm excited and enthusiastic about the application of AI in medical imaging"},{"startTime":2883.338,"endTime":2891.405,"body":"and in oncology in general."},{"startTime":2892.425,"endTime":2897.865,"body":"Um, the revolution alluded to in the title is underway, but it's not yet very"},{"startTime":2897.865,"endTime":2900.405,"body":"tangible in its impact in the clinic."},{"startTime":2901.505,"endTime":2906.142,"body":"So we can see that these new tools that are now beginning to be developed"},{"startTime":2906.142,"endTime":2908.925,"body":"and deployed look like they're having real clinical benefit."},{"startTime":2910.865,"endTime":2911.085,"body":"Now."},{"startTime":2911.145,"endTime":2914.845,"body":"The speed and the scope of the AI developments are startling at the moment."},{"startTime":2915.075,"endTime":2918.605,"body":"However, the approach to medical research is necessarily much more cautious."},{"startTime":2919.905,"endTime":2925.167,"body":"Robust evidence from trials and practical experience from the early deployments will need to accumulate"},{"startTime":2925.167,"endTime":2928.325,"body":"before we're ready to rely on totally novel methods."},{"startTime":2929.385,"endTime":2933.815,"body":"But if the new generation of AI technologies is able to deliver anything like their"},{"startTime":2933.815,"endTime":2938.245,"body":"promise, then we might just see some entirely new approaches to some aspects of healthcare."},{"startTime":2940.465,"endTime":2942.125,"body":"So thanks very much indeed."},{"startTime":2942.425,"endTime":2948.275,"body":"Um, I'm hoping that some people in addition to kind of questions have also got"},{"startTime":2948.275,"endTime":2954.125,"body":"comments and, and statements and, and answers to some of those things that I've raised."},{"startTime":2954.545,"endTime":2957.005,"body":"Um, and yeah, so thank you."},{"startTime":2961.775,"endTime":2965.91,"body":"Thank you so much Richard for a really wonderful thought provoking talk and for sharing"},{"startTime":2965.91,"endTime":2967.565,"body":"your children's really wonderful art \u003claugh\u003e."},{"startTime":2967.565,"endTime":2967.805,"body":"Thank you."},{"startTime":2967.825,"endTime":2968.565,"body":"It was very, very good."},{"startTime":2969.105,"endTime":2972.873,"body":"Uh, we do have some questions on the tablet already, so if you're happy for"},{"startTime":2972.873,"endTime":2973.124,"body":"me to ask the first one and to the, the question who asked this, I'm"},{"startTime":2973.124,"endTime":2979.405,"body":"going to elaborate on it a little bit if that's okay."},{"startTime":2980.105,"endTime":2984.085,"body":"So the, the question is, can this detect lobular breast cancer?"},{"startTime":2984.865,"endTime":2990.305,"body":"And how I would like to expand upon it is when we have the publications"},{"startTime":2990.305,"endTime":2990.668,"body":"which looked at the concordance between the AI and obviously the human read, was there"},{"startTime":2990.668,"endTime":2998.285,"body":"any difference between the subtypes, lobular, ductal,"},{"startTime":2999.905,"endTime":3000.205,"body":"Et cetera?"},{"startTime":3000.205,"endTime":3002.725,"body":"Yeah, that's a, that's a really good and important question."},{"startTime":3003.025,"endTime":3008.975,"body":"Um, and uh, people are looking into to that, but I think we don't know"},{"startTime":3008.975,"endTime":3010.165,"body":"enough about it."},{"startTime":3011.705,"endTime":3018.165,"body":"My observation is that the same sort of cancers that the humans and the models"},{"startTime":3018.165,"endTime":3023.765,"body":"work in a more similar way than I perhaps expected that they did."},{"startTime":3023.945,"endTime":3029.69,"body":"But the, there is more overlap between what AI models detect in terms of breast"},{"startTime":3029.69,"endTime":3031.605,"body":"cancer compared to two humans."},{"startTime":3032.185,"endTime":3038.146,"body":"So that gives an opportunity that perhaps between the two, between the human reader and"},{"startTime":3038.146,"endTime":3038.544,"body":"the AI reader, you can detect more cancers altogether if you then arbitrate the disagreements"},{"startTime":3038.544,"endTime":3038.941,"body":"correctly or somehow you make the right decision about which ones you're gonna act on"},{"startTime":3038.941,"endTime":3054.045,"body":"because you still get lots of false alarms as well."},{"startTime":3054.745,"endTime":3061.337,"body":"Um, I think that that question about lobular cancer is really interesting because, uh, it's"},{"startTime":3061.337,"endTime":3061.776,"body":"my view that really what we should be doing with these AI systems is, is"},{"startTime":3061.776,"endTime":3071.885,"body":"using massive volumes of training data, population level volumes."},{"startTime":3072.225,"endTime":3076.605,"body":"And then we can start to train an AI system to look"},{"startTime":3077.835,"endTime":3082.89,"body":"Just, uh, at the most dangerous cancers, uh, to look at, to classify things not"},{"startTime":3082.89,"endTime":3083.227,"body":"as cancer or not cancer, but you know, it thinks this is a, a grade"},{"startTime":3083.227,"endTime":3083.564,"body":"two lobular cancer or, or it thinks this is, yeah, this is a, an eight"},{"startTime":3083.564,"endTime":3094.685,"body":"millimeter possible high grade cancer."},{"startTime":3095.445,"endTime":3100.445,"body":"'cause then we could start to use approaches where actually if it looks like if"},{"startTime":3100.445,"endTime":3100.778,"body":"it's falling into that category of a much more dangerous cancer, then we accept many"},{"startTime":3100.778,"endTime":3107.445,"body":"more false positives in that situation."},{"startTime":3107.695,"endTime":3112.847,"body":"False alarms, if we're gonna catch the ones that are most likely to lead to"},{"startTime":3112.847,"endTime":3113.191,"body":"bad outcomes compared to, yeah, if we try and train for much more indolent cancers"},{"startTime":3113.191,"endTime":3113.534,"body":"that actually may not matter, may not need diagnosing at all because they might just"},{"startTime":3113.534,"endTime":3126.245,"body":"stay there for the rest of the patient's life."},{"startTime":3126.765,"endTime":3129.005,"body":"'cause we know that, that that situation occurs as well."},{"startTime":3129.745,"endTime":3134.307,"body":"So, but to, to be able to train for that, you need really large volumes"},{"startTime":3134.307,"endTime":3137.045,"body":"of cancers in, in your, in your training set."},{"startTime":3137.585,"endTime":3144.458,"body":"Um, and I think that's part of why kind of it, part of why the"},{"startTime":3144.458,"endTime":3144.916,"body":"question about kind of population level support and participation in, in the, the technologies that"},{"startTime":3144.916,"endTime":3153.165,"body":"are developed is important."},{"startTime":3153.895,"endTime":3154.685,"body":"Thank you very much."},{"startTime":3155.505,"endTime":3161.452,"body":"Do you think that privacy laws and new regulations will accelerate or slow AI's integration"},{"startTime":3161.452,"endTime":3162.245,"body":"into medicine?"},{"startTime":3166.865,"endTime":3174.565,"body":"Oh, well I think it's vital that we, we protect the privacy of patients."},{"startTime":3174.585,"endTime":3179.992,"body":"Can you, can you imagine if you go and have your mammogram and then it's"},{"startTime":3179.992,"endTime":3180.353,"body":"you, you start being advertised brows of a particular size that's just right or, you"},{"startTime":3180.353,"endTime":3188.645,"body":"know, uh, it it is absolutely vital, isn't it?"},{"startTime":3188.785,"endTime":3197.885,"body":"Um, this area of regulation of, of privacy is just evolving all the time."},{"startTime":3198.425,"endTime":3203.654,"body":"Um, I think people quite often do feel that it impedes things, but I I,"},{"startTime":3203.654,"endTime":3204.003,"body":"my view is that it is just so important to get it right because we"},{"startTime":3204.003,"endTime":3211.325,"body":"need everyone to trust what's going on."},{"startTime":3211.945,"endTime":3216.598,"body":"So if it takes us a little bit longer, i, I, you know, it it"},{"startTime":3216.598,"endTime":3216.908,"body":"needs to, it needs to be, it needs to be trusted and everyone needs to"},{"startTime":3216.908,"endTime":3225.285,"body":"be happy with it so that we can make the best of it."},{"startTime":3225.625,"endTime":3227.005,"body":"Hi, um, thank you very much."},{"startTime":3227.005,"endTime":3227.565,"body":"It's great talk."},{"startTime":3227.585,"endTime":3232.17,"body":"Um, so I'm coming from more the AI side rather than, um, radiography, I dunno"},{"startTime":3232.17,"endTime":3234.005,"body":"anything about radiography to be honest."},{"startTime":3234.105,"endTime":3239.525,"body":"Um, so I was just more curious about the training or data pipelines."},{"startTime":3239.525,"endTime":3241.845,"body":"If we think about the deep learning models for now."},{"startTime":3242.425,"endTime":3245.045,"body":"Um, like is this data publicly available?"},{"startTime":3245.115,"endTime":3248.085,"body":"Like how large are the data sets that we're talking about?"},{"startTime":3248.825,"endTime":3251.965,"body":"Um, yeah, just, just curious about that training process."},{"startTime":3252.555,"endTime":3260.525,"body":"Yeah, so there are some publicly available data sets, um, particularly from America actually."},{"startTime":3260.605,"endTime":3267.557,"body":"'cause I think their research, uh, infrastructure demands that in certain situations you make things"},{"startTime":3267.557,"endTime":3268.485,"body":"publicly available."},{"startTime":3270.265,"endTime":3274.685,"body":"The one, so I've been involved with A-C-I-U-K project called Optimum."},{"startTime":3275.025,"endTime":3281.334,"body":"Um, and that's started as a kind of physics project looking at what factors in"},{"startTime":3281.334,"endTime":3281.755,"body":"mammograms made cancers, uh, either be permissible or, you know, the, the kind of physics"},{"startTime":3281.755,"endTime":3288.485,"body":"of mammography."},{"startTime":3288.905,"endTime":3294.287,"body":"Um, and that changed into, uh, a big data set that's been useful for training"},{"startTime":3294.287,"endTime":3295.005,"body":"AI systems."},{"startTime":3295.945,"endTime":3305.005,"body":"So, uh, that, that, um, database originally came from three hospitals."},{"startTime":3305.045,"endTime":3307.325,"body":"I think it's, it's now expanding to six or seven."},{"startTime":3307.985,"endTime":3315.342,"body":"Um, and you can apply to, it's not publicly available, but it's the public can"},{"startTime":3315.342,"endTime":3318.285,"body":"apply for, uh, access to it."},{"startTime":3318.625,"endTime":3323.995,"body":"So, uh, academics, um, uh, can have access to it for a small fee and"},{"startTime":3323.995,"endTime":3324.354,"body":"commercial companies can have access to it for a large fee is kind of my"},{"startTime":3324.354,"endTime":3329.725,"body":"understanding."},{"startTime":3330.185,"endTime":3332.485,"body":"Um, and that's, that's run by CR uk."},{"startTime":3332.985,"endTime":3337.2,"body":"Um, and it's the guys at the Royal Surry County Hospital who, who, uh, have"},{"startTime":3337.2,"endTime":3338.325,"body":"kind of built that."},{"startTime":3339.205,"endTime":3345.745,"body":"I wonder, COVID demonstrated the NHS, uh, provided a national basis for research and a"},{"startTime":3345.745,"endTime":3347.925,"body":"lot went on in Southampton."},{"startTime":3348.545,"endTime":3354.416,"body":"Um, and there was a huge commitment amongst that local population to make their, uh,"},{"startTime":3354.416,"endTime":3356.765,"body":"data to participate in that research."},{"startTime":3357.305,"endTime":3363.597,"body":"Uh, I just wondered as regards your topical, uh, uh, inquiry, whether an elderly cohort"},{"startTime":3363.597,"endTime":3364.016,"body":"concentrating on them would be, would enable you to get more results, um, more effective"},{"startTime":3364.016,"endTime":3373.245,"body":"results, uh, rather than a population-wide, uh, study."},{"startTime":3373.905,"endTime":3379.845,"body":"Um, obviously older people are gonna exhibit those signs of, uh, uh, cancer more new."},{"startTime":3379.915,"endTime":3381.485,"body":"There's gonna be more of them literally."},{"startTime":3381.705,"endTime":3381.925,"body":"Yes."},{"startTime":3382.225,"endTime":3389.285,"body":"Um, I mean cer certainly that that could be the case for your training cases."},{"startTime":3389.905,"endTime":3394.717,"body":"You can have a kind of biased population, for example, over sample, but have many"},{"startTime":3394.717,"endTime":3397.605,"body":"more cancers in that population than than you'd see."},{"startTime":3398.105,"endTime":3403.639,"body":"Um, and so you, you could train a system using kind of unbalanced population like"},{"startTime":3403.639,"endTime":3404.007,"body":"that so long as you then also have sets of data that represent what you"},{"startTime":3404.007,"endTime":3404.376,"body":"want it to achieve for the kind of the validation bits and then crucially the"},{"startTime":3404.376,"endTime":3415.445,"body":"test sets."},{"startTime":3415.905,"endTime":3417.805,"body":"So yes, I think so."},{"startTime":3418.105,"endTime":3418.325,"body":"So"},{"startTime":3418.715,"endTime":3424.12,"body":"This is a question I'm really interested in your thoughts on and indeed those in"},{"startTime":3424.12,"endTime":3428.085,"body":"the room for rare cancers which prove challenging for clinical trials."},{"startTime":3429.025,"endTime":3432.43,"body":"How can we overcome the lack of data on them to train the AI with"},{"startTime":3432.43,"endTime":3433.565,"body":"and have that test data?"},{"startTime":3436.515,"endTime":3437.685,"body":"Yeah, that's a really good question."},{"startTime":3438.665,"endTime":3444.973,"body":"Um, and, and these systems, I mean, why breast cancer has been attractive is 'cause"},{"startTime":3444.973,"endTime":3445.393,"body":"it's got really loaded, it's common cancer with loads of data and our, our brains"},{"startTime":3445.393,"endTime":3453.805,"body":"are very good at being adaptable."},{"startTime":3454.625,"endTime":3459.725,"body":"And so when we learn to do one thing, we can kind of apply that."},{"startTime":3461.455,"endTime":3463.115,"body":"So yeah, that's a, that's a real problem."},{"startTime":3463.535,"endTime":3468.365,"body":"It might be less of a problem for this new generation of AI systems that,"},{"startTime":3468.365,"endTime":3468.687,"body":"that they've, they, they claim kind of zero short learning where and, and where it"},{"startTime":3468.687,"endTime":3469.009,"body":"can do things that you just wouldn't expect it to be able to do that"},{"startTime":3469.009,"endTime":3479.635,"body":"it hasn't seen in training."},{"startTime":3479.815,"endTime":3484.675,"body":"So who knows that that might, that might help address it."},{"startTime":3485.725,"endTime":3487.765,"body":"I think it's definitely an area where there'll be focus."},{"startTime":3488.975,"endTime":3489.925,"body":"Where there say again, sorry."},{"startTime":3490.005,"endTime":3493.485,"body":"I say I think it'll be an area where there'll be focus considering the high"},{"startTime":3493.485,"endTime":3494.645,"body":"unmet need for rare cancers."},{"startTime":3494.645,"endTime":3494.805,"body":"Yeah."},{"startTime":3495.385,"endTime":3498.125,"body":"Um, my question was very simple, I think."},{"startTime":3498.305,"endTime":3505.28,"body":"So, um, how do you think genes or silent genes would be used for detecting"},{"startTime":3505.28,"endTime":3511.325,"body":"cancer from genomics in a person who's exhibiting cancer from your radiography screen?"},{"startTime":3511.465,"endTime":3515.859,"body":"So how would you use the gene or a silent gene to confirm cancer is"},{"startTime":3515.859,"endTime":3516.445,"body":"my question?"},{"startTime":3517.505,"endTime":3522.485,"body":"How would you detect cancer from a silent gene or a gene from your radiography?"},{"startTime":3523.945,"endTime":3530.361,"body":"Um, I mean, both imaging and genomics have a, a really strong potential to kind"},{"startTime":3530.361,"endTime":3531.645,"body":"of work together."},{"startTime":3531.885,"endTime":3539.493,"body":"I mean, the kind of genomic medicine, these kind of liquid biopsy type techniques might"},{"startTime":3539.493,"endTime":3544.565,"body":"well replace the need for image-based breast screening altogether ultimately."},{"startTime":3545.025,"endTime":3549.701,"body":"Um, so yeah, I don't, I don't know how they'll go together, but I can"},{"startTime":3549.701,"endTime":3550.013,"body":"imagine that, that, that working out the nuances of what's important about a cancer, both"},{"startTime":3550.013,"endTime":3550.325,"body":"to be able to diagnose it accurately in the first place and then to know"},{"startTime":3550.325,"endTime":3562.485,"body":"how to treat it, that that could work very nicely together."},{"startTime":3562.585,"endTime":3562.805,"body":"But"},{"startTime":3563.795,"endTime":3568.304,"body":"What are the biggest obstacles to get this adopted on a wide scale in the"},{"startTime":3568.304,"endTime":3568.605,"body":"NHS"},{"startTime":3573.705,"endTime":3574.835,"body":"Kind of logistics?"},{"startTime":3574.955,"endTime":3583.205,"body":"I think, um, we've got quite an obsolete, uh, information system that breast imaging runs"},{"startTime":3583.205,"endTime":3583.755,"body":"on."},{"startTime":3584.135,"endTime":3590.6,"body":"And there is, there, there is a, a call out currently for running a big"},{"startTime":3590.6,"endTime":3592.755,"body":"trial of AI in screening."},{"startTime":3593.135,"endTime":3599.266,"body":"So that hasn't been concluded that that funding call, and it certainly hasn't started that"},{"startTime":3599.266,"endTime":3599.675,"body":"trial."},{"startTime":3600.535,"endTime":3606.835,"body":"But I think the, the hope is from the people who wrote that call, that"},{"startTime":3606.835,"endTime":3607.255,"body":"that trial will also kind of force a system to be developed that can then"},{"startTime":3607.255,"endTime":3618.595,"body":"be kind of turned into the next generation of, of screening information system."},{"startTime":3618.965,"endTime":3620.235,"body":"Thank you very much again, Richard."},{"startTime":3620.235,"endTime":3620.715,"body":"Well thank you."},{"startTime":3620.715,"endTime":3624.106,"body":"Thank you Nikal thank you to Novartis for sponsoring it and we very much looking"},{"startTime":3624.106,"endTime":3625.915,"body":"forward to seeing you at the next lecture."},{"startTime":3625.965,"endTime":3626.395,"body":"Thank you."},{"startTime":3626.495,"endTime":3626.915,"body":"Thanks."}]}