[CVPR 2019] Le Lu, Executive Director @ PAII Inc. Bethesda Research Lab
Dr., lu lu is the executive director of pi bethesda. Research lab hai, is a tech, driven company focusing, on smart city technologies. Intelligent. Education, programs and technical. Solution, on healthcare, imaging and clinical informatics, prior. To pi he worked in the National Institute of Health clinical, Center his, research focuses on improving, the modern medical image. Understanding and semantic parsing to, fit into revolutionary. Clinical. Workflow, practices. We. Thank you for joining us today and welcome, of course the cvpr. Yeah. I'm looking forward to our discussion, so. You received your PhD in computer vision and, machine learning from, John Hopkins University in, 2006. Can. You tell us more about your PhD research. My. Research, is about general, configuration like, I. Before. That I'd be a party or some research in cultivation. 90s. Like so basically, 1990. To 2000, this. Is the you know geometry people. Doing, like a projective, geometry. 3d. Reconstruction, for. Ten years. Then. From, 2001. The, safest. Also. I will kind of co-education five, my first officer at that time you get a flavor of deep learning deep learning in. General can, be considered, a special case of political. Mass or statistical, massive it's just one it's just differential, in, essence right. And, we're, at cvpr. And we're three days into a very, intense program and, a lot of information, that's being discussed, here what. Have you seen you, know what's exciting you about the research that you're seeing that the, folks are discussing we. Had a. Kind. Of couple like a medical. Shopping. Hydrant, well, workshop and also I was ready to be worth talk at another workshop, so, I was most folks my, own workshop which is the fully liquor medical. Imaging yes. Ace. Here are we running this portion called medical comparisons, from. Fifteen actually each time we saw full house we're, early morning to learn that. Today, I saw, a couple of things I think very, interesting other. Panelists, our research, topic they held weekly morning so this, morning is a few chocolate Ernie and. Area. Share for the serious year so, I did a seat. And. I think, we've the most dependents, ever CVP, or the highest attendance is. This year so it's great to see that the program. And the conference is developing. So fast maintaining, such relevant seat as well yeah, so, NIH, is the primary government. Medical research agency, in the US and. You worked there for over six years and ever received, the NIH clinic CEO, award on research, excellence and. Patient. Care impact, and, it seems you've primarily focused on applying machine, learning and deep learning in, image, processing especially. For, radiology image, and image segmentation. Could. You summarize your major, achievement. And learnings, during your work at the NIH, I. Would. Know just complete with everything to the, work I. Did. I'm Raj I spent, six years before, and I spent six years in the industry I realized. The limitation, by just. Working in the industry right, being a little bit distant from the physician. So. I decided, to pursue a career. After. The government GSA in, the hospital, so basically do some research nobody house can. Do it less, for. Us we're unique we are the largest. Really. Focus. Translational. Hospital. On panic, not just in us the. Biggest, planning. So I work in lucky now hopefully serve. At. Least or doctors aware, of famous doctor and very, engaged in research, the, research in clinical means, like an oompa have, a new cure well. New guidelines. About treating certain, disease. For. Better patient, outcomes so, I was, so. Is there for like five six years. Failure. At the employee. Well, actually you as a special volunteer, we are one of, the early group after, many, imaging, on, top of deep learning for. Many images, we. Started at at 2013. As, a time like a lot of people don't believe you, can use but. If learning format, in the media because you have in that area has 1 million plus images, 1.2. Million images. We have labels, precisely, in. Essence for many people. So. We make the, progress.
Energy. Progress actually, when when face down like an energy leave no detection in. Are integrally the whirring for mile marker pens efficient and. We. Improve, the permanence. Relatively. From like a. 55%. For say we, moved it to 84, percent so. That's one thing with you how. Deep learning can really be applied to a linear they said like a 150. Patient, for example and, it. Can work amazing, well compared with previous work, you, know nothing, about mentioned the defining about some me into we, do everything you have been doing well using traditional answer, but, it, more, means you, so certainly, very important for me not so well, previously. You, know you can solve it so. That's the one major strategies, we making a certain. You know like. I'm you need an, annual. Work, for. I come. Very good detective America. Texas and also. Our work last. Year we cut off every, year a. Word. It's kind of you, know the most important, very much failures in essence so, this is a cat especially. Kinds, of provisional, diagnosis. And detection, but. Starting from there is totally realized, this. Is important, Portland is a former mostly study you, can many on hospital, many on the startup company they actually doing this they're headed to indicator. Community. The detection. But. My eyes a, better, problem facade or. More. Platform. Who saw who is a higher. Urgency. To solve is actually precision, medicine precision. Medicine. In. Out in a remote. Field it. Has direct. Higher. Impact, for. Our. Practice. No. I swear. In the street, for in fact they have since so, basically you. Go to the clinical practice they, look at the turn of work of trained, doctors. Actually know the. Another necessary means the path of treatment guideline. For the patient uh best way. Or technicals, in certain disease. Like. 20 minutes 25 minutes for one, page and. This only, the presser feet but wait you can get the edge of human so. They do not have much help from the computer or from computing, right so, for example, well. One typical thing you do is you measure your. Science of watoomb. So. In. Siri is, federal, matter the warning okay. But, this is not possible people. Doctor we don't have time to pin, how the voxels, are tumor in 3d and, computer, one in essence, and, also you need to have this one over time because. Often. A patient is, kind, of under certain treatment, she, is feeling working on that so right. Now the Fargo is now the idea is a worry menu, so you eyeball, the. Site when his roof I can foresee a sly, new things yes has, the highest. Exchange, special, attention learning. You member the longest, the diameter, now, you let's, call Allah exists, now, you measure, the short axis which is orthogonal to long, axis so this axis. Is basically, in. The. Informatics. Recorded. Into the patient record right, rather, than how big for large the tumor it is but, suppose you should matter if anyone of Matt research right so, over, time when the next time if even, the same village in doing that and. You. Recommend. Were not precise. Very. Go down to, easily doctoral, work without being abstract, reasoning which, in which machine cannot do it by today women we, don't know. So. This kind of measurement is, as simple as is something. Michigan. To where well I'm, human, and make an error it.
States Is. This tumor tumor. Grow. 10%. Yeah. That means coffee different things in a clinical indication. This. Is called this, is one application University. Medicine, for. Tumor, method and for. Tumor Manliness critical is simple it was a simple thing but this, you see you can make, a big difference, this. Is. Positive. Imaging, precision. Medicine he won't matter in say we're, accurately, where, it consistently, over time this. Is one thing we're working on we. So. Certain, automatically. Like. A. 40,000. Deaths every, year the US all, compared. To cancer he's worried that anyone funny is almost no cure because too late is where difficult to find so, we're. You, know we can help you some yummy developers, model. The warm, and shave and around. That formal, walk walk so great was something like a manufacturer. Finishing, understand, nope, you, know the Machine friends tell physiologist. This is, horrible. How, have, you mean happy no. Medicine. Yeah and all of this I mean thank you for those explanations, as. Well I mean it's certainly we, know that this is in our future and this is going to be more informing. How medical, research is is, being wrong as well as we move forward so. When, we think about that and we think about precision, medicine and, other, types of research as well we know that there are a lot of restrictions to, applying black box methodologies. Like deep learning to clinical, trials, how. Did the how, do these. Restrictions, reflect, your research. Would say it's. Hard to argue differences. Being a black box yes, it's not totally expendable. Or. The. Architecture, is to contact me too many parameters. But. I do now feel this really, big. Issue right because. The. Thing is you you, will not be able to because. As my model before. Then. Also. They're. Not very human yeah the. New second feature. From, millions. Of feature that's, a free money how boosting work I think, the ticket or something. Maybe. A better way it's. Easier, to be expending I agree, we need research, along this line to, make it more expendable. Wait. Don't. A sad-ass model. Itself, may, be extremely, more you, have billions of Frankfurt Hagen you understand, what's wrong, the. Thing is, we. Can make it we. Can make it a more manageable, in. To know what we learn the, object, we're. Not just at. So. For example something. Female. Worker yeah. So you feel learning, something, you tell human. Situation, this is the vernacular all. Money, going to marry how, can you, verify you. Have no way to verify. Black. Box I just a common issue is you, cannot, use it the. Caucus. Because. It's. Not work out. Casey is actually. Learning something you. Can you have to tell the future why you think. Morgan. So, you have. I. Told, you how to murder I can't, be defined away if you have the, petition. Incomplete. Instead, of making the penalty say you know why you, can make a decision, found Y prime. Or y1. Y2 what you buy, so. You learn maybe the. Texture features. Certain. Shape feature it's right, explain. Is, there this, little task. Human. Time worth it, human. Know if the machines are making mistakes right so after you have oh this minor feature this will, work, over minor features, from.
There You, need to can have a Cadillac saying. You. Can make a decision, say for example to integrate. All, this meta, feature into, a fantasy is Marie Gaddafi, from. There the million I can, contribute, you, can come, in. This. Matter feature. Tasks, so, you for, so narrow to learn a lot of sub tasks. Then. You put an earlier, decision. Making, mechanism. Depending. On this work of all features, and their particles, will be expendable. Paco, is very simple, right now human use finishes. 30. Minutes to make a decision which, is right. How, we doing, something. Providing, more, information, then. Now much. More information, than now does, it make a better decision, with. 15, minutes so. -. Yeah, you lower the time. And. You can working on something nicer and combine. Them together. By. Using less. Composition. Time we. Can make a better decision much. Better efficient outcome. For. The society, yeah, yeah. I mean there's always the commercialization. Of course. Yeah. At. Some point I'm, so moving, on from that and when we we, look at your current role you. Join PI as executive, director about. A year ago what. Are you and your team currently working on okay. First I'm gonna explain you to be like that what sky. Research. She copied. Founded. By one, of the largest teachers. So, juicy sure come they come including, like a finance. Investment colas. Is a fact. And lotto like, how scary disabilities. Yeah like many. People each or so. But, one thing I truly believe is you. If you look in the medicine you know the, us with divided into, three. Parts three part is my patients. Are, wider, house, care providers, the other is appear. Usually. Compass. Providers. And. Others. And. When, you think from PI's perspective. Then looking. At these problems like, what what work are you what. Work are you and your team doing that to address that I think the big turkey is that actually I. I. Know, what, I'm trying to do you, know I'm trying to you know receive managing, and, how. What's, the we can offer resistant to more patient okay, sure right this. Is. The future is. For. The star for a. Report. Hey I 100. 2016. Yeah my. Other way sir he, won't a part of it, from. There they were saying is you. Seen us you. Know wishing harder will came from like a kind. Of like a like a passive permanent a mmm-hmm, okay sinker why, because, they're, posting sure it goes be sure is comfy and also, their house care providers the, hospital. And the interns are the same so, they are motivation. About, providing. Better house care reduce. The cost because, yes no medium there's no application about, to, party they fight each other, in. Essence yeah the only customer. Is patient, so, the they hope they can you. Know I hope there, could be many in this kind of hospitals, or medical difference. Hospital. So, you NASA's. You'll. Be fighting for better, efficiency. Lower. Cost to. Guarantee. Patient, better vision, better.
The. Manage your house they see if, someone get a disease, so they want to curate that as, very as possible, as you know I saw it also with better, because. That, makes, sense you caught me right, make, sense I hope I won't be that's the future so, I think the PI is helping. He's, part of the formula, how. Come, the component, 2 makes it happen and so the parent, company of the, PI is, one of the biggest infant, company, and we, have. The fishing. You. Know they decide this. Also the way to go so instead, of being a passive. Payer. For. More like a more. Active, patient. Management house. Manufacturing. Company, we. Get to the premium or come to your house in, Essex so. So. Yeah that's that's our, function so what we, want to doing the similar things I was. An energy but. Different. Right now it's all different because you know I was, a lot of primary physicians, I can, the full spectrum about, the. Patients. In. The hospital, I, use, walking in a video Department. Americans. Image. But. You know don't. Really control the patient it directed manage the patient they're, more like a consultant, Department am in the hospital so, the primary position where initially. Can also help, a so we, don't know what's going on so the, other needs. Patient, you taking this you can. Report. Back in so. That. Can resolve. The past right. Right, now we mostly. Working with the primary physicians, to, know the what's the puzzle so, we have more. Interest, in. Integrative. Medicine so, means like a you meaning is the one involving. Part of its you know I'm fine, with train but. We should look in the whole picture of the patient house, and the, whole spectrum all patient. Management not just. Meeting on from that as well because we've been talking a lot about research, we've been talking about the application on hospitals, the relationships, with the doctor, I mean this is this is where, we're going is a certainly, the future is very bright in. Terms of those improvements and collaboration. And partnerships, so. I want, to end with this final question to you around many AI researchers. At the beginning believe, burning can disrupt everything however. Deep learning is data hungry and performance, is strongly correlated with, the amounts of available training data in, your perspective what, are the boundaries and limits of deep table deep learning and I know we've covered some of this before as well but if you were to summarize, this where where do you see today, and, then where do you see where we can lead to so we're we're gonna move forward to. This. Is actually, a wonderful question, so first of all we film into learning valley we, do, not learn the data of IRA Henry way in the sense since I proposed, is when, you learn a lot of things among. Excel. Regularly, we. Can talk to each other like a we can model the. Religions. Different. Maybe. Maybe, you only need a few mapping. From X to Y you. Can get a problem solved this is my mathematical. Answer so I'll give you an hour sorry you know sorry I also, talked to the hospital. To definitions, looking. At the problem is physician. Will never. Be. Employed, to create a simple. In the. Same. Way has the abdominal, jerk, or something right Felicia will not really know it for labeling. Images, between, computers, I said.
Take, The, example. The example. The, basically is the, term of finishing fewer puzzles, say as this is author answer, me and, he actually tried to read, the patient prior what happened before why. Is this is the first thing meeting is the second time he was this follow-up, imaging now, a look at the images probably you know kind of previous, images right so, you have images, questions. And, also what he learned from the medical school. From. His fellowship, right so, this is kind of ontology. Knowledge. A base kind. Of thing well he or she three. Pieces of even being together. Yeah. That's, what you do three, machines already. Viewed. A lot of informations, to. Try doing the dealing work if you think. Describing, what. This image right is, a it's a three times before but. That contains, one human procession is about this image right forty, percent what we do is we're trying to. Field. No. Matter to, our last, year plus a few year we have six, we. Have six, CPR. Papers and. 1wc, reversal, was all about that yes we, travel to leverage. Clinicians. Clinical. Annotations, this. Is something that produces a routine, work. But it's all the information, which have no crack down decoding. What's your sale, and, using, that as a label to. Learn. Images. Instead, of everything. You read. Analogy, is saying you, know you, might have a small house hospital, right we, have a meeting, Fisher. If. You arises. Here. For, the past one year after year will, you people to to, do that just by creating a new design label the same way 100. Doctors here we have nothing the cost. It. Cost a lot another. You do have time do it so. The best way is already. There so, we need a new master, of all, the worried dying what women, would say we you, know I, I cry to the word from this clinical. Center CEO it was how, we need somewhere big they said Helen Cesare and that's. What you see here ago, my, last you care what, they said how more. Than 10,000 via studies, ultimate. Consumers, and think. About, dr. Marek, there's, also has some how can be maybe, sir either. The. Tag. And the, labels would not produce, about Amazon, Mechanical Turk, we, cannot do it was, mined, from doctors, in Port Arthur information, and, even. Which a visual grounding. So. You haven't seen your master the, argument kind self. Annotated, so. And, this also learning program yeah. We human do that that's why we can learn from of your. Email. Pinyin that's the future all we have to do this kind of normal matter. I. Maintain always, the human centered approaches. Always. See. Thank you so much for your time today this is a great conversation thank. You.