 I lead the cognitive computing and data sciences lab at Cognizant. Cognizant is a MNC doing service across three lines of business and we were originally founded in 1994. As far as the last Fortune 500 list is concerned we are at 195, we employ around 2.5 lakh associates globally and I am here to talk about our experiences in the medical space and the influence of deep learning and AI in medical space. Today's agenda is in four parts. First we will briefly talk about medical imaging itself kind of level setting the ground and we will talk about trends and challenges that are there in the industry and then common solutions that are used to overcome some of these challenges and finally we will share our experiences. There is some of course some time for Q&A so my request is to hold on your Q&A if you have any. So I understand this is a mixed audience so people are from different groups. The next set of slides is to make sure we are all on the same page of course assuming we are on the same book for all practical purposes. So some of you might have questions on what is so great about medical imaging right this is just another technology supporting a stream of science right. So if you look back prior to technology supporting healthcare let us assume a patient or a subject comes and complains that he has got a physical pain to a doctor unless and until the bone is sticking out of his hand and visibly people can see what is happening there is no way for the doctor to say that he has got a fracture right. So they used to go into something called experimental surgery so what that means is technically cut open and see. Now the patient would actually have lot more pain than the previous pain right so it is going to be lot more confusing so you can imagine the amount of wrong diagnosis the amount of life threatening events that could happen possibly because of that. From there medical imaging has brought us to a point where even the brain functions can be non intrusively mapped and figured out right. So medical imaging and the importance of that is so significant in the scheme of things that the first award Nobel prize in physics was given to the guy who invented the x-rays just in case you guys already know that. So for the next 30 minutes we will actually use a simple definition so that we are all on the same page again. So we will use this definition thanks to Wikipedia for telling us what is medical imaging. In simple terms medical imaging is the technique and process of creating visual representations of the interior of the body for clinical purposes right. So there are two types of medical imaging one is diagnostic the other one is non diagnostic names are intuitive very simply put non diagnostic imaging is used not directly to help a physician but indirectly some of the examples could be common examples are image guided surgery or brain computer interface right. Subsequent slides that I am going to talk about is primarily on diagnostic images and things generated based on diagnostic images. So when we think of medical imaging we think of x-rays, MRI scans and CT scans right. If we have personal experiences it goes beyond one or two of that right but that is not the case. Medical imaging means images of tissue, fundus, ultrasound, ECG, EEG, endoscope the list goes on and on and on right. You can see that they are very very varied set of images it requires special skills to understand and read. I remember when I first held my x-rays right side left right and doctor came down and said that is not the way to hold this is the way to hold ensure all of you have your own experiences on that side. So that how complex these domain is. So where do physicians really start? So diagnostic images are grouped into five categories, radiology, microscopy, photography, graphics and others. X-rays, CT scans and MRI scans which are used for screening muscle and bone abnormalities fall under the radiology umbrella, ECGs, EEGs are grouped under the graph. Later in the session we will touch on histopathology and retinal which are part of microscopy and photography respectively. The wide variety and the volume of different images that are being generated in the space has led to a sub stream which brings us to a very important topic in medical imaging called computer aided deduction or diagnostics. Again sticking to the original theme I want to make sure we all talk about the same set of definitions. So we will again go back to Vicky and borrow this definition for this presentation. When we say at an abstract level they are the same but if you go into the nuances they are slightly different for the sake of this conversation let us assume they are the same. Computer aided deduction abbreviated as CDE, computer aided diagnostics abbreviated as CDX both are systems that look at medical imaging point out abnormalities and assist the physician or the clinician to do diagnosis and subsequent prognosis. Some of these systems have become so advanced that they just do not call out abnormalities anymore. They actually go down to tell you whether a disease is there or not, whether there is a likelihood of disease happening or not. They also go down to the level of grading some of these diseases for you. So they have that much of processing capabilities now. Some of them go to the ultimate step of even giving recommendations kind of replacing some of the very rudimentary analysis that a physician would do. So moving on this part of the presentation will briefly look at the evolution of these medical imaging systems and the images themselves and how deep learning fits into all this and then we will look at state of the art and what is trending in that space and of course we will look at the new set of blockers and challenges these technologies bring to the table. This graph by the way goes all the way back to 1660 till a few years ago. The first official description of photoacoustics was published by Bell back in 1880 150 years ago believe it or not and it took 100 years from that point for us to actually have an MRI scan available. So it took that much of years for things to evolve despite the basic discovery it took that much of time. You should not be surprised if I tell you the high definition image of a living tissue was taken less than a decade back that is how complex the whole space is. So what was the real reason why the evolution took long lack of light sources. So people understand light so we all understood physics right when you throw light it gets observed or reflected human body is not like that a very simple example is when you throw light on your body the same tissue that observes today might reflect the quantity of absorption and quantity of reflection varies right. So that is how complex the whole system is. So it took time for scientists to understand the whole space and it evolved. So the source the deduction technology the data acquisition and the processing capabilities where the real limitations for us to kind of move or the reason for the 100 years to get to where we are right. Again now medical imaging have come a long way now we actually are looking at digital radiology we are looking at keyhole surgery we are even talking about augmented reality when it comes to doing some of the surgeries right. We have clearly shifted from physicians pretty much doing everything manually which is the era from all the way up to the 80s to systems based on heuristics helping or assisting the physicians which is the mid 80s to the modern day deep neural networks and CNNs which are actually helping them go past and helping them to diagnosis and prognosis beyond that point. So this shift is primarily driven by the fact that people wanted to actually get better care improved clinical outcomes and they wanted lot more accurate outcomes people do not want to do trial and error on themselves. So that is the history part of it right. So where are we now in very simple terms systems have started out doing humans that is the reality of the situation I have thrown in three examples here couple of them from Stanford one from Cornell the first one that we talk about uses deep neural net looks at skin images and predicts whether skin cancer exist or not in par with the dermatologist the Cornell one uses neural net actually looks at X-rays and predicts fracture in level with the radiologist the last one from Stanford actually looks at chest X-rays and detects pneumonia. So the reason why I stuck to academic thing keeping in spirit with the conference this is publicly available the data the approach the outcome. So you guys can look at it critique it reuse it extend on that right but the nonacademic side of things the big guys IBM's Microsoft's Google's they have all been making claims on these and not much of it is publicly available for us but I am hoping all of that would eventually become publicly available at some point in time the spirit of answering the question towards the end I would actually request you to hold on to that question if that question is not answered towards the end we will revisit that and answer that question. So let us actually look at a different perspective right we saw that things have started out doing the doctors again since that question prompted I want to make a disclaimer here that does not mean all the disciplines of medical imaging have actually started out performing the doctors no that is not what I am saying here what I am saying is there are selective areas in which systems have started performing in part if not above the humans right. So with that said let us actually look at the analyst perspective of how medical imaging is shaping up right analysts are projecting hundreds of billions of dollars to be coming in to the system the next five years right market leaders are significantly investing in devices in the medical imaging space that use AI and ML and deep learning right some of them are around personal medicine patient data management and all around that right the major players are actively scouting for buyouts we keep reading mergers acquisitions all around I am sure all of you are aware of deep mind from Google IBM acquired explorers and vital PTC acquired coal light right clearly analyst and the industry is not looking at AI and medical imaging as a competition to the physician but helping him make the right decision quicker right that is how the industry is looking at it. So let us look at a different perspective now we looked at the academic side of the world academy has always been the forefront or the vanguard of technology research whether we like it or not right. So these are the recent ones that are got published so if you can see there is a clear spike in the amount of publications patents that are being done on this this numbers by the way is specific to deep learning in the medical imaging space and there is a reason why you see that from 2012 or 2013 I will get to that in the later slides and there is a clear bigger number if you were to actually add the allied. So this does not include the classic ML this does not include all the other branches of AI this is pure deep learning artifacts that are in the medical image space. So if you take a closer look at the publications themselves right while classic ML are still being used almost one third of the artifacts are using neural nets right. So if you take a closer look at the neural nets themselves you can clearly see that CNN is the most prominent or the one that is gaining significance right and while others are still there they are still being used CNN has made a significant impact when it comes to medical imaging okay. So what are the current challenges? So does it mean this has solved all the problems that physicians had before of course some of the problems have gone away does it mean it will be too naive of us to think that there are no other problems technically speaking this systems have introduced new set of problems to us right. So what are those? It is a labeled desert as far as the medical imaging is concerned what does that mean there are images everywhere so CAD systems are throwing images more frequently now and none of them are labeled very few set is labeled. So now you do not know which to use which not to use to train your model. So that dilemma is kind of going to haunt us. So the next one is the data overload so clinicians are in data overload now there are systems everywhere generating a lot of data now what technically we wanted to do was to simplify the physicians life by generating this data but now what we have done is we have overdone that. So now there is lot more data for the physician to look at. So technically this has become a little counterproductive. So the next one is digital silos in the interest to gain the market share most of the industry what they have started doing they have started throwing in new systems. So these systems have their own proprietary way of capturing these images storing them processing them because they have their own way of looking at it they do not talk with each other. Now the doctor does not necessarily have to look at lot more sources to get the information he has to also understand different systems and these systems do not talk with each other. So he has to look at different systems to kind of come to his final conclusion. The next one of course is the compliance and the governance. So with the social era and increasing demand for instant gratification of course we want to actually have more compliance and governance FTPR I am sure all of you are aware of GDPR and how do we ensure that these systems store these information securely and share only to the legitimate stakeholders. And the last one which is in my opinion the most important one is the trust. How do humans really start trusting decisions made by these systems? How do we make them transparent? How do we make them less bias? How do we remove the data bias or the algorithm bias so that people can really look at what is happening? How do we make the whole thing transparent? We do not have answers to all of them, we have answers to some of them. Some of them still continue to be riddles that is for the research community to kind of come up with more solutions down the line. So in this section we will refresh some very fundamental items about deep learning neural nets and in the context of digital imaging again and we will also look at some possible solutions to the challenges that we saw in the earlier slide. Okay, quick peek into the history of how we got here. Initially before the pre-deep learning era there were two distinct styles. One was ML with feature input and followed by the ML with image input itself. In the pre-DL era we would actually do sequential applications which did a lot of low level pixel processing followed by mathematical model which was actually computed using a simple heuristic rule set and it was a point solution that we created. So a simple pixel operation would be like edge deduction, line deduction and stuff like that. Mathematical models would be line fitting, curve fitting, items around that. So since the 90s, supervised techniques came into picture that started pushing the whole neural net in a different direction. The real push happened when back propagation came into picture. That was in the 80s, late 80s, early 90s. So prior to back propagation it was only feed forward. So basically you take your input, do your analysis, get the output, if the output doesn't match what you want, you manually go back, tweak things, then send it again and hope it works this time. With back propagation the error aspect, how far is your prediction, how far is your classification from the actuals, is taken, fed back into the system and that made the system lot more better. Then we had this whole problem of vanishing gradients. So what does that mean? The feedback that was coming in was not sufficient, the system was not learning. So what did that mean? That actually meant we couldn't have deep nets, features were not fully getting identified, your models were not converging. So compute power was not sufficient to get to what you really wanted to get to in the time that you had. Then something really happened in 2012, unsupervised pre-training, Alex net coming into picture opened the floodgates. So that is why in the previous graph in 2013 and forward you see a lot of research work happening because that in very ways minimized the vanishing gradient problem. So from then on people could have a deeper neural net, a wider neural net and it could be lot more efficient when it comes to getting the outcome it wanted. So after 2012 the deep learning and the CNN hasn't looked back at all, in many ways it is only forward and we have done a lot of stuff after the point. Let's actually talk about some very very fundamental things. Assuming all of you understand that computers don't look at images the way we look at, we look at images they look at numbers. So a pattern recognition, in a pattern recognition feature is nothing but a unique measurable property that uniquely identifies a particular image. Very simple terms, if you show the image of a human, nose, eyes, mouth, all that forms a set of features. The aspect of identifying these features is called feature engineering. So we have some basic tasks like classification and localization. Detection is nothing but taking a new image input, trying to find out which of these buckets it belongs to. In this case it's a sheep image, it's been classified as a sheep image. What is localization? You identify your region of interest, ROI for short. So we often draw a boundary, here our interest is the sheep, so we have drawn a boundary around that. So the next one is object deduction. So in the image if we have more than one set of objects that are of interest then we call them out. In this specific case we actually have three sheeps, if we had a dog and a cat that is of interest you would have actually seen the bounding box around that as well. And the next is segmentation, right? Segmentation in many ways is logical grouping of the items that you see in the image. There are two types, instance and semantic. Semantic is just clubbing all of them together. This is declassifying them inside of the instances themselves. Hopefully you guys are still with me, I haven't made you sleep yet, I haven't seen anybody use the two-step policy that was described in the earlier thing, good. So what is the basic difference between classic ML and the deep learning as we go forward? So the feature engineering or the feature extraction that we spoke about in classic ML happens manually. So you get hold of a subject matter expert, you get hold of an engineer, you give him an image of a human and say okay, now you tell me what is the feature. The SME would actually say two eyes, nose and mouth makes this image distinct, the engineer would code that and then the engineer would decide whether this is a classification problem based on the task, pick the algorithm, send the input through it and get the output. Again in the interest of earlier, I like your question, we will definitely get to that question if it is not answered towards then, yes, the short answer is yes, we will get to it. So in deep learning, what we do is the manual intervention has reduced, I am not going to say it is eliminated, it is reduced. You pick the architecture and the architecture decides what feature to pick and what task to perform. So the amount of manual intervention has dropped, if you guys were there in the morning speech, you would have heard him talk about auto ML which is the next level of deep learning where people are not even expected to do this, you just give an annotated sample input, everything else is done for you behind the scenes, I am hoping excellent, okay, did I do something? Hopefully that was not the case, okay, keeping up with the difference again. So classic ML and deep learning, okay, before I go further, let me just quickly highlight what is anomaly, that is a new term I introduced in this slide. In very simple terms, it is to identify the outliers. So you train a system to spot things that are different from what it is being trained on. So based on which school of deep learning you are from, you might look at it as classification or you might actually say no, no, that is a different discipline altogether on the other end of the spectrum, okay. So that is a different topic, we will discuss later. So for all practical purposes, these are the different tasks. Classical ML has all the algorithms available to do them and deep learning has the same set of algorithms to do them, right. So if I were to actually take a second step at this, what really is happening is classical ML improvements on the algorithms have slowed down, right, because deep learning algorithms are becoming more accurate, predictable and faster. So in essence what I am trying to tell you guys is while classical ML is still there and it is still happening, the pace at which things are coming out now have slowed down. You would have seen that in the graph that we showed earlier, right. So with deep learning that is not the case. So all of it is moving towards a neural net specifically CNN in case of digital imaging and digital image processing space. So I have been talking about CNN, CNN, CNN all, let me just tell you guys what CNN is. You guys already know, bear with me for the next few minutes, right. So what is CNN? CNN is nothing but another artificial neural network that is specifically put together to spot patterns. What helps them to spot the patterns? They actually have a layer called convolutional layer which help them to find these patterns. So I do not want to tell you guys that CNNs do not actually have fully connected layers. They have convolutional layers also. They are like any other artificial neural net but the only additional thing is the convolutional layer, right. So what is the convolutional layer? Convolutional layer is nothing but a set of filters. So for simplicity filter you can assume as a simple two dimensional array. Let us assume there is a 3 by 3 matrix. So we initialize it with some random numbers. Once we get an input image, this 3 by 3 filter is slid over the input image all the way through till there is no other image to browse. This process is called convolution and that is why we have the name convolutional network. So I apologize for all the statisticians in the room. So what happens when it slides over each of these pixels group is a dot product of this matrix is done with the source matrix and that becomes another number. And we create a separate matrix at that point and that is fed into subsequent layers. Hopefully you guys have with me so far. Since there are multiple matrix and a lot of multiplication just like any other neural net there is a pooling layer. For those who do not know what pooling layer it is just the way to kind of abstract the dimensionality so that it is easier to handle subsequently with minimal loss. Of course there are different ways to do that the most popular one is max pooling which means you just pick the top one in that move and move on. So in a convolutional setup the initial filters are for identifying simple patterns. It could be lines, curves and stuff like that. The deeper the filters you add they gain the capability to do lot more complex patterns. It can actually be to identify the eyes, nose, it can even identify the human face. So how deep you go kind of decides the quality of information that you are going to gain from the input. So the final layers are the only connected layers in a CNN. So they do the aggregation and the final task which is classification or whatever that you want to do. The initial layers are to get the patterns and the extractions from the input image. Hopefully you guys are with me so far. Now have we actually now that we understand how CNN works the problems that we had earlier right the unlabeled data has that gone away no we still have that problem how do we overcome that problem right so that we can use this CNN to efficiently do the tasks that it want the network to do right. So two of the most commonly used approaches one is called transfer learning very simply put you take the knowledge from a pre-trained neural net and use that to do your tasks for a different domain right. So the philosophy there we have we have all seen teachers students right teachers actually go through the process of learning and then telling us and students actually kind of grab that. So it is very similar to that from a philosophical perspective. So what does this really mean? So this is an example from one of our own experiments. So as I said earlier in the CNN description the last layers are the aggregation layers. So we kind of remove them and we replace them or swap them with the way we want to actually do the subsequent tasks we use the earlier convolutional layers and that knowledge is transferred over okay. So in very simple terms you do not have to go tune your parameters one more time because this model is already built it can actually do a decent job of predicting the outcome so there is no waste of time you can leverage all that and move on to what is more important to you to actually get the output right. A word of caution there is you probably want to pay attention to what is the base model that you pick. The reason why I say that is let us assume you want to do a classification of human specimens right man versus cow versus maybe an exotic bird it is okay to actually pick it from the image net database right the model that is built on top. But if you want to actually do a phase deduction you are better off picking up something from VGG than some other else right. So try to see if you can pick the base model closer to the problem statement that you have already solved or you are expecting to solve. So the next option is to do data argumentation to overcome the data problem that we actually have. So very simply put you take whatever data is available to you apply something on top of it turn it flip it rotate it and do all that from a neural net perspective it would be looked at as a new set of data for it to train right it is as simple as that. So some of the commonly used techniques are geometric methods photometric methods and adversarial methods of course the GANs by themselves is a different topic altogether. The geometric methods are flipping cropping scaling rotation I am sure all of you guys are familiar with this these are standard stuff that we do on most of the imaging techniques. Photometrics is where we play with the color filters we apply additional histograms on top of it or remove noise add noise and stuff like that and from a GAN perspective just to give you guys for those who are having trouble what is GAN it is a network which actually has two parts to it the best analogy I can think of is a faker and a cop put together in the same one. So basically you take an input the faker part of the network kind of morphs it and creates a new image from it the checker part tries to identify if it is able to find out if it is fake or not so that is that is the that is the setup again that deserves a full session by itself so I do not want to kind of go into that detail but from a technique perspective that is a technique people use for data augmentation to overcome the data problem. So what else is happening when it comes to the trends part of deep learning right from a object identification perspective so there is a concept of domain transfer GAN where you take image inputs from one domain and then reuse that for another domain for training the models right the other popular ones that are being in the academia being heavily looked at is one shot learning and zero shot learning as the name suggests they want to actually use minimum number of inputs to build your model both are variants of transfer learning in in practical purposes the the last one no no discussions in CNN would be complete if we don't mention capsule network right so very simply put it's it's a neural net within a neural net so you you actually have neurons that activate on an individual basis right these are some of the ones that are happening right now on the digital image again a disclaimer there are a lot more work happening on the DL on the non-digital imaging as well but these are the ones that directly affect the digital imaging and the medical space okay so the this next set of slides I'm going to talk about is the the couple of experiences that we have had right the first one that I'm going to talk about is cancer detection that we actually worked on for breast cancer and the second one that I'm going to talk about is the diabetic retinopathy using fundus images both were built using deep learning networks and this helped us to appreciate the domain and the technology much much better prior to us starting it right breast cancer is the most common form of cancer and woman it is so common that one in eight is actually globally diagnosed with it I believe and it's the second leading cause of death in woman simply because 60% of the deduction I believe happens in the final stages right we pick this focus area almost 12 16 months back and our intent was to look at digital whole slide images to help pathologists classify these images as normal benign malignant and grade them as in situ are invasive right in situ technically means the tumor is within the organ that that originated that and invasive means it's already spread and we are in the final stages and it is life-threatening from a process perspective what really happens is there's a physical examination done and the doctor looks for unusual lumps and if the subject is less than 30 years most often it's not a problem but if the subject is more than 30 years the first order of business is to get a mammogram or an ultrasound done and if those two kind of direct to abnormalities then what is done is a biopsy in very simple terms a piece of the tissue or the lump is taken out a slice of it is taken a coloring dye is put on top of it and put us under the slide and that slide is called the whole slide right and that image is digitized so all of us would have seen slides right they are this big but when you actually put them under the microscope and zoom it 400x they actually become much bigger than the screen so imagine a pathologist looking through this big of an image finding out abnormalities right so it is a tedious process and it is very very labor intensive there is a fair chance a lot of it is going to be missed so that's why this is being looked at as the last resort so what we said when our idea was to see how we can help them so the way we kind of went about doing it is our thing started after the digital imaging part of it we took the whole slide image broke that into patches manageable smaller chunks of patches we looked for normal benign and malignant within that and when we found a pattern we put them together right post that what we did was we actually did segmentation we clearly called out the different nuclei within the image we colored them or highlighted them in a way that the pathologist can understand then as part of the quantitative analysis we started giving the count of mitotic and a atypic nuclei so so that the the pathologist can take those numbers and compare that with the upper and lower boundary and eventually do the grading of whether something is cancerous or not right so the whole model was built on transfer learning we took the base model from an image net trained image and then we built on top of it we fine-tuned it as I said earlier we lost the last aggregation layer we added our own layer so that you can do what we wanted that to do moving on the the next case study I am going to talk about is diabetic retinopathy case study apparently this is also a leading cause of blindness and if you don't detect this earlier enough it can lead to permanent blindness in humans and people who are diabetic are extremely prone to getting diabetic retinopathy right so just like any other medical setup the patient to doctor ratios very skewed here as well right so when when when there was an NGO Karnataka based NGO they reached out to us and they said a we want you guys to help us we realized the gravity of the situation in a developed country itself the ratio of patients to doctors is very skewed in a developing country like India it is so skewed that the patient does not get to see the doctor in a very long time right first time he comes in and sees the doctor and the next time he's going to get a chance to come and see is going to be way way too long by the time he might end up becoming blind as well right so when these NGO came and reached out and told us and explained us the problem we were we were kind of very excited to get on board and see how we can actually help right so we got on board around 24 months back and we kind of put a solution together for them and now we are in the pilot phase with them and we are actually doing pilots as we speak so from a solution perspective right we we built the model originally from a 35000 base image okay and in that process we started off with CPU slowly graduated to GPU now we actually have an NVIDIA DGX-1 for the processing capability power again based on the best practice when we started we did something called binarization what that means is technically lose the color information in the image bring it to gray scale and we actually did normalization technically what that means is change the size of the input image right irrespective of whatever the source images you change the size we we were fully convinced because these were the best practices suggested at that time that we will get the best of outputs and the output was not that great and then upon reflecting we realized that we are losing valuable information when we do this so we started retaining the source image size we started retaining the color information and that made a lot of difference when it comes to the output prediction right we actually used yes okay so okay let me give you some more insights into this so the way it is practically done is a person whose subject puts his eye on a device and a camera operator takes a photo so there's a flash of image a light flash is done in your retina that is reflected back which is observed and that is stored in the digital image right so just like how you take a photo camera of yourself the same thing is done but inside the eye so the the glare there actually means overexposure or an unwanted exposure to a specific area right so not getting into too much of detail the pupil size varies from geography to geography and the amount of light that you relay also has to be the intensity of the light that you relay have to be controlled based on that right so this is so complex that even if the device is the same if the operator is the same if the patient is the same if you take two shots the two images will not be the same that is how complex the whole setup is right and of course camera artifacts like there might be a mosquito sitting right in front or a simple dust or speck of a dust mosquitoes too much the guy can't see a mosquito then he's blind already right so there could be a speck of dust there right so all that adds to the camera artifact okay so again we wanted to do transfer learning in this the learning was not to do transfer learning in all the cases we built these model from from scratch after going through the journey of course we extensively used data argumentation in this because of the fact that we couldn't get enough samples right okay moving on so what was our learning from all this right invest in your data set see most often people assume that deep learning programming is all about coding in our experience it is less than one-fifth a significant chunk of time has to be spent in getting the right data okay so more does not mean better okay I'll tell you a practical example we had so much of left-eye image samples which were non-DR that if you give an image to us we were prone to tell that you don't have DR because that's the samples we had right then we realized that that is that is bad and then we started applying many data argumentation techniques and started hunting for quality images so if you have a five-class classification that you want to do your training image sample size should be evenly distributed across right make sure that you have quality data for you to build your models on the next one model is not a black box okay there are sessions happening in this conference and all around which is talking about explainable AI and understanding why certain things are being given as outcomes by the different models again that's an evolving space I'm not saying everything that the machine does is explainable at this point but there is some amount of explainability that can be derived from what is being done right now and the last one is in keeping spirit with that of the conference and similar conferences like this being engineers we tend to start writing everything from scratch right so I have that habit myself so I think the other one is not right or I don't actually believe in what that is right so we start writing don't do that start off by borrowing ideas and models that are already there it's a two-way street by the way right so once you think something is working I would suggest you to contribute back as well right so the two case studies that we spoke about both are in pilot and we are hoping to see lot more of them coming out and that actually brings me to the end of this session so I first of all want to before I take the questions thank the odc for having me here thank you guys for listening patiently so now going back to the two specific questions that you had are they still open do you actually okay no problem so can you repeat your questions I'll see if I can answer them okay so the straightforward answer is we are not trying to replace the doctor with this and I don't think that is going to happen right there are very very rudimentary things let's take the DR example that we do the actual screening process the photo taking process the screening process of somebody has diabetic retinopathy or not the initial screening is not done by doctors it is done by technicians and they are short in number so the ratio is so skewed that 1 is to 100 plus is probably the number we are looking at so these systems are not trying to take away jobs from these technicians or doctors these systems are trying to address a market space which was left unaddressed if these systems were not there that's that's the way I would rather look at it and doctors can do lot more efficient work so very simply put if you were to actually look at cardio CT scan I believe there are thousands of images a doctor has to look at comprehend all of it and then look at what is of interest to the patient which is very very rudimentary stuff to look at abnormalities which can be done by the machine so the doctor can have an efficient 8 hour window where he can still perform the same job at even better outcome compared to what would have otherwise taken 16 17 hours he can actually have a proper work life balance if I were to pull it that way so that's where all these systems that's where I believe these systems are going towards you had a second question I see few more hands but so the answer is if we have labeled data then there's nothing like that excellent but who will label it the same doctors who are sitting and addressing the patients have to label it right that's where the it's a it's a predicament you understand that right so it's a chicken and egg problem so the doctor will he actually be looking at the patient or coming and actually labeling the data set for you right so that is where if you have labeled data excellent if you don't have a labeled data we actually have techniques which will build on top of existing labeled data or we actually use unsupervised techniques to see if we can proceed further so labeling is an excellent option good path if we don't have we still have to figure out alternative ways otherwise we are going to get into another winter session like I would like to add on to the answer which you gave for that person will doctors replay I mean I will deep learning or this data science replace doctors see I think even if it replaces it we should encourage it because in countries like India and Africa there's not many lot of people don't get the medical facilities and stuff as you mentioned the ratio of doctor to patient is very huge especially in sub Saharan Africa and India companies so this is one area where we need data science to splurge and you need that advancement really and one more thing is it is not going to replace doctors I I mean all your case studies were related to breast cancer and I but I've done a lot of research in neuroscience specifically so I've been reading a lot of literature in neuroscience and the at the advancement of data science into that so those case studies were quite interesting so those are areas where doctors cannot make an advancement without the help of data science so there's particularly a case study if you're interested in the audience is interested this particular case study by MIT okay it's called optogenetics where it is like finding the accurate area in your brain so for disorders like schizophrenia and all the psychological disorders the I mean the constraint which they have is they want to find the accurate area in your brain where the neural pathways are inactive so for schizophrenia patients there are several areas in their brain which is inactive compared to the patients who are the normal patient will have you know neural connections in those areas so their challenge is that so with the help of data science now there is a lot of they've tested it out in rats so they are able to like find out the exact regions and brain where which are inactive and stuff with the principle we agree yeah yeah yeah and there's also research which is happening in Princeton lab it's out in YouTube so all of you can check it out so there's a lot of interesting case studies in neuroscience which is actually happening so I think this is one area where all of us should support the advancement and yeah even if I mean by the way are you involved in neuroscience research and cognizant yes we actually are exploring at least my team we are looking at eG images we are looking at ECG images we are trying to understand I'm not a liberty to disclose everything and that are different teams working on different items well we are assisting some of them yes and short answer is yes is the team does the team have an equal amount of domain experts or it's purely a technical team or again I'm not a liberty to discuss certain other items outside my domain so to we can probably take it offline and we can probably I saw a few more hands go there I don't know how are we doing on time hello basically yeah so you showed the state of the artwork going in Stanford Cornell I don't know where this voice is coming from it's coming from here yeah so hi I am Ashish so basically in the industry also they have R&D teams which are working on this topic yes deep learning and so my question is like when how far we are from commercialization basically in the breast cancer area also detection of that basically what has been going for a long time correct but still I don't see a lot of products coming out of it so so there are different reasons for that just to extend on your earlier part of your question I did not purposefully talk about the the non-academic side simply because some of this are maintained as trade secrets they don't want to disclose lot of it is not disclosed except for one or two random papers published by Google or Microsoft here and there I want to keep to the spirit of the conference so the openness is what we want to actually promote so that's why I kept that to answer your first part of the question yes there's lot of work happening and the reason for that is I'm going to say a different thing here hopefully you guys can still follow there's the whole concept of Confucius matrix right false positives true positives false negatives true negatives what is the implication of something kind of drives then some of these solutions become commercial so let's assume an outcome falsefully identifies something as malignant the implication to that is less slightly less compared to if the same thing says that you are not malignant when it is malignant right so that kind of drives some of this commercialization aspect so to on the positive note FDA has approved the diabetic retinopathy screening device now the device can actually do a screening and tell you whether you are retina or you are subjected to diabetic retinopathy or not right so we are getting there but a lot of this is simply because the supporting technology is also evolving and the trust aspect that I was talking about there is less of transparency which is why there's a delay so we are moving in the positive direction just to answer your question and if you want to give me a timeline I have no idea your guess is as good as mine one last question any inputs on validation how you validate that brought a point about FDA approval and all so how it is taken in medical imaging is a lot more complicated so we have worked with lot of partners and support systems so the way things are looked at is also very different so software is no more a software if it actually has a direct implication to a human right so if it is it can be classified as a device right a device by itself if it doesn't actually do a diagnostic thing then the set of certification that is required for it is very different and exactly when it comes to Europe it is different when it comes to Asia it's different when it comes to India it is very very different right the only mature market I would say is US which is where FDA is driving some of this they have significant amount of norms laid down for devices software alone device plus software so that I would say has taken them anywhere from two to three decades now to get to where they are and they themselves have come out and agree exactly they have come out and agreed that this by itself cannot be the final one right so they are selectively doing a lot of this so in that I forgot the actual question I don't know if I answered it I have an interesting point of view on that right so from a software perspective the way we do testing and validation is very different from how the medical domain is looking at some of this so if you go tell them that I did a test run for two hours a clinician would look at you and say are you kidding me I have actually been looking at the solution for 10 years and I still am not confident that this is going to work so they are worlds apart to be very simply put so in our experience of taking some of this to a practical field what we realize this an inclusive approach where we kind of bring them on board and keep telling them it is it is end of the day they do not know how to trust what is coming out of the system so transparency exactly so it is evolving space is what I would actually some last question please thanks Mr. Balaji for your nice interesting talk I want to ask you how the data augmentation is done when we are transferring the domain okay so they are overlapping so when you actually say a simple example I can think of is the five streams that we talked about in medical imaging right so we have seen a lot of radiology images taken and applied to photograph so we take those apply one of the three techniques that we discussed transform it into something else and move it on so again adversarial networks by themselves is a big discipline and again I hope I kind of one and to add on do you think that in future the radiology should be expert should be an expert acid stands as we discussed earlier there are some expertise already built by the AI so radiologist at some places can be replaced as it speaks but again it is the future prediction is anybody's guess at this point maybe