How Machine Learning is Transforming Radiology | Chad McClennan | TEDxNorthwesternU

you artificial intelligence we're all curious about it some of you might even be a little bit concerned what can it do what impact is it gonna have in our lifetime artificial intelligence as a term was dates back to 1956 it was originally coined by John McCarthy and the traces back even farther to the work of famous mathematician Alan Turing I'm not going to give you a history lesson I'm just going to tell you that I believe it's about time that this is going to transform modern healthcare I'd like to start with a single moment of discovery the story begins 2007 a team of engineers sitting in drab office cubicles in New Jersey they're working on computer code for defense contractor doesn't sound very glamorous and it wasn't sound powerful absolutely you see these engineers were working on face recognition so that our US army could catch criminals around the world what if I told you that this same technology that was used to catch bad guys could be adapted to catch bad lesions bad nodules or what we also call cancer you see these engineers their work involved using sophisticated biometric data radar technology and machine learning to interpret images infuse that information with other information so that the Department of Defense could quickly and accurately identify people and objects sometimes even inside buildings without ever having to go inside sound sci-fi this was real then why stop there they thought we could zoom in see this person why stop at their face what if we add additional information that allow us to look inside their whole body they were using radio waves to identify people could they use other waves like sound waves to see what was going on inside what uses sound waves to look inside our bodies ultrasound yes we use sound to see you know those adorable little black-and-white pictures that parents stick on their refrigerators it's a little girl sorry that's a little boy as it turns out we're already using ultrasound to monitor flow of our heart detect plaque in our arteries and look for malignant tumors these guys knew they were on to something they asked themselves what if what if instead of finding bad guys which I defined suspicious tissue instead of making radar and radio waves work better which could use sound and sound waves work better not just at finding something but actually capable of telling a physician what that something actually is distinguishing bad from good without ever having to go inside after six years of data collection experimentation the model started working and working well today this smart technology is poised to be a tool to help make lots of physicians better how because they use sophisticated biometric data in models that are the fingertips of physicians but who am I to be telling you this not a physician not a scientist not an engineer I'm a father and a husband with a daughter and wife I've had some pretty serious health scares they're fine now but as a family we spent our fair share of times staring at state-of-the-art medical equipment with technologists physicians with few answers sometimes wrong answers guess is at best confounded many of you might have the same issues and concerns worried about a loved one terrified of the word malignant we all know that an improper diagnosis or treatment path can tear lives apart how does the word exploratory surgery sound to you these guys who said what if are now those engineers using machine learning to battle cancer Who am I I'm just their CEO I'm doing everything in my power to help them succeed you might be wondering is this going to be a complicated dissertation on machine learning absolutely not it's the practical application it's about getting information and data to make decisions how do you make your decisions you ask around you seek guidance from experts you phone a friend you compile information the more accurate and reliable information the better so we begin with sound waves sound waves tell a story it responds differently to the underlying tissue and cellular structure it presents itself in the form of images these images are pixels and they tell that story just the way sonar does underwater because our bodies are mostly that water we look at the end of these images and the data that lies within them these images our patterns and pixels pixels can actually be used to predict lab results machines and their models can begin to learn these patterns much the way humans learn patterns and physicians are trained to they'll see things that are invisible to the human eye and impossible to ever correlate a physician might see seven ten patterns in an image to justify a diagnosis these models can be trained to see thousands some so subtle that we would never ever as a human ever be able to pick up these patterns will allow us to see and predict pathology results we can classify this pattern as benign or malignant and put us on the path to an accurate diagnosis come quickly we must do something now or wait no trouble found false alarm what blows my mind is the more good data we give these models they start to learn for themselves they're given answers as well as image data the answers that we're talking about here is ground truth ground truth this is the definitive answer of what something is so when the Machine looks at the image its seen it before this ground truth is sourced from millions of biopsies that are performed annually more is better the more ground truth we have and associate with these images the better these systems can learn this is the collective experience of thousands of experts using a machine and a model to analyze this image becomes the equivalent of doing a digital biopsy here's a fun fact the speed at which these machines now can process information is about a million times faster than our already amazing human brains another way what the machine can learn to do in about a week which a team of us about 20,000 years so we have to get over ourselves machines have become smarter than we are if we define smart as being correct a lot and really fast we have to stop Lepik letting ego and fear get in the way of intelligence you see our eyes can play tricks on us we see things and our minds process those images based on colors and contours and shapes and shadows but it's shocking how often we can see one thing and believe that it's something completely different sheep dog a mop or both Chihuahuas blueberry muffins which and just a brick wall right you can't unsee the cigar right in the middle once I tell you that it's there you see our eyes playing tricks on us and being fooled can be the same as misinterpreting medical image and ultrasound CT it leads to errors misdiagnosis and worse amiss cancer all too often this is a function of human bias myriad factors can contribute to bias can be time of day level of fatigue scientific journal you just read or the patient that the physician saw earlier in the day or their very last patient what's the likelihood that this next patient also has a chronic disease you see experts disagree with one another all the time we've conducted multiple studies that indicate that experts will disagree with themselves up to a third of the time and get this physician will disagree with him or herself up to 20 percent of the time you see we've given the same images to board-certified physicians twice separated a month apart a month later they're asked again render their decision one out of five times they come back disagreeing with their prior diagnosis but equally as confident you see this is the soft underbelly of Medicine according to study 8 your study conducted at Johns Hopkins medical errors are the third leading cause of death 10 percent of all deaths surpassing respiratory disease we can do better than this we will be better than this when we take advantage of these technological advances that are at our disposal what was once doubt can now be trust machine learning provides an objective way unbiased way to analyze images we started proving this out in breast cancer why it's the leading cause of cancer-related death in women around the globe an ultrasound is oftentimes the best and only viable modality those cute black and white pictures to determine whether there's something malignant and it can vary widely based on the skill of the sonographer operating the probe or the skill and experience of the physician reading the exam false positives and false negatives missed cancers are way too common sad reality is a story of misdiagnosis happens every day here in Chicago across the country around the world but imagine if you had AI embedded in the software to serve as a second set of eyes that software can look at that lesion the way a physician would in medicine that second opinion could be the difference between life and death engineers around the world are integrating that ground truth the machine learning and the models to service that second set of eyes and it's available on a moment's notice in real-time or on-demand see the past two years our researchers have been proving out the viability of this second opinion model this extra set of eyes this software embedded with AI in partnership with physicians from Columbia University Memorial Sloan Kettering and others under the guidance of the US Food and Drug Administration that 15 physicians recently analyzed 900 cases with and without AI assistants 27,000 case reads guess what we have found using an AI to assist the physicians outperformed themselves sometimes by a very large margin they were catching up to six more cancers per hundred while simultaneously the cases they recommended for biopsy dropped some cases in half catching more cancers earlier while simultaneously reducing false positive positive biopsies this appends conventional wisdom conventional wisdom is the only way to catch more cancer perform our biopsies what's this mean to you and I your friends your loved ones the second opinion these models can catch more cancer earlier while still treatable while reducing over treatment miss diagnosis the risk and anxiety that goes along with it you see the future doesn't arrive all at once it's early yet but this is so promising and it's so logical I'm predicting that physician adoption will be far more rapid that most expect they've known this is coming for years but we've entered a new era today Google can process billions of web pages in seconds it's transforming industries it's gonna impact healthcare imagine if we would do with medical images what Google does with these pages and analyze billions of CT scans MRIs ultrasounds and direct a treatment path the minute an abnormality is found see more and more physicians realize that with AI assisting the guesswork diminishes and that frees them up for more complex problems and to spend more time with their patients this is just progress this is a race and we're in it we now the AI horsepower the ability to put this technology into imaging machines and make them easier to use we've literally connected it to the mouse the cursor as a physician analyzes an image until it's in use every day people are getting care that could be better some cases life-altering ly better what happens next it's up to us the pace of this technology is a function of our laws and our ability to acquire data performance of these systems rely heavily on the access and accuracy of data the time has come to think big and be fearless like those engineers who said what if what if instead of just using this detective detective cancer you could actually inject and use a nanobot to inject a cure right at the cellular level what if we knew so much about the disease inside you we could map it to yet DNA and stop it before its Rhett's what if instead of just on adults this was children or fetuses before they're born with diseases or disability so I'm not proposing that AI is going to replace humanity I'm arguing strongly that it will help humanity we need you you can help you don't need to be a data scientist you don't need to be a physician this is for all of you here your family battled cancer and for those thanks to AI that'll never have to thank you [Applause]

Channel: TEDx Talks

Views: 6,541

Rating: 4.6831684 out of 5

Keywords: TEDxTalks, English, Health, Medicine, Technology

Id: OvIJkSmIAuY

Length: 18min 12sec (1092 seconds)

Published: Thu Jun 13 2019