Founder Stories: Andrew Beck, Co-Founder and CEO of PathAI

Good afternoon. My name is Immaculata D'Adivo, and I'm one of the co-directors for science programming here at Harvard Radcliffe Institute. I would like to welcome you all to our first program in the Harvard Radcliffe Institute Science Lecture Series on artificial intelligence. Today's topic is artificial intelligence for pathology with our featured speaker, Dr. Andy Beck. The Harvard Radcliffe Institute began its multidisciplinary explanation— exploration of artificial intelligence in October with its science symposium, Decoding AI: The Science, Policies, Applications, and Ethics of Artificial Intelligence. You can view the video recording of that event in its entirety on our website. As a cancer epidemiologist, I can certainly attest to the critical need pathology plays in the diagnosis of disease and the development and implementation of tissue-based prognostic and predictive biomarkers. In this talk, Dr. Andrew Beck will discuss the potential for recent advances in artificial intelligence to significantly advance the accuracy, reliability, and predictiveness of pathology with applications in both clinical research and practice. But before we start our program, I'd like to take a moment to acknowledge the members of the Radcliffe Institute Leadership Society and the annual donors who are watching this afternoon. In particular, I'd like to recognize the Mary E. and Clara Z. Costanza Fund for Science, Ellen and Melanie Mason, and W. and David W. Niemiec Fund for Science, which are supporting today's Science lecture. Your generosity keeps Radcliffe programming free and open to the public, and we thank you. And now I'd like to briefly introduce my friend and colleague, Dr. Andy Beck. Dr. Beck developed one of the first machine learning-based systems for cancer pathology. He has been certified by the American Board of Pathology in anatomic pathology and molecular genetic pathology. Dr. Beck was on the faculty of Harvard Medical School in the Department of Pathology at Beth Israel Deaconess Medical Center. He has published over 110 papers in the fields of cancer biology, cancer pathology, and bio— biomedical informatics. He co-founded PathAI to develop cutting-edge technology that promises to better diagnose and treat some of today's most challenging diseases. After his presentation, Dr. Beck will be joined in a moderated discussion with another friend and colleague, Dr. Akila Visanthawan. Dr. Visanthawan is the chair of the Department of Radiation Oncology and professor of radiation oncology at the Johns Hopkins School of Medicine. Once they conclude their discussions, Dr. Visanthawan will then turn to moderate the audience Q&A. I would like to encourage those watching to use the Q&A feature on Zoom to submit their questions at any time during the program. Since we anticipate a lot of questions, we ask that you keep them short so that we can answer as many as possible. It is now with great pleasure I pass the virtual floor to Dr. Andy Beck.

Great. Well, thank you very much for the introduction, and it is my pleasure to speak today. I also look forward to taking part in question and answer after the session. And I'll be speaking now and for the next 20 to 25 minutes or so on artificial intelligence for pathology. And to provide some background and context, I want to give an overview of how pathology plays a fundamental role in the patient journey. And to start with a case from Dr. Joanne Elmore, who is a cancer epidemiologist at UCLA, who underwent a skin biopsy, as indicated there on the right. To evaluate whether a growth on her skin might be a benign nevus or a malignant melanoma. This really represents a typical patient encounter with pathology where a patient will have symptoms, undergo biopsy, send the sample to the pathologist where the pathologist is asked to answer some fundamental and core questions like, does the patient have cancer? How advanced is the disease? And which treatments are most likely to be effective. So it's an area where it's really critical that patients have access to the most accurate and reproducible diagnoses, uh, to then help guide treatment decisions as well as to inform them of the prognosis of disease. And Joanne Elmore is not only a patient, she's also, as I mentioned, a cancer epidemiologist. And she published in the British Medical Journal about 4 years ago a study called Pathologist Diagnoses of Invasive Melanoma and Melanocytic Proliferations Observer Accuracy and Reproducibility Study. And in this study, they did a very simple but powerful analysis where they took a set of skin biopsies. This is one indicated here. And these are the routine type of biopsies that are seen by doctors all over the world. Here's an example here on the top left. And what this purple and pink image is, it's a microscopic image of a patient's biopsy. And this would be viewed by a pathologist under the microscope. And this is one actual example from her study. So this is a single biopsy. And for her study, she sent this to 36 pathologists. And for each pathologist, asked, what is the interpretation of this image? And for this particular image, this in her study received 18 different diagnoses from these 36 different pathologists. And these 18 diagnoses are categorized into these 5 different risk categories from the study. Study as indicated here on the right. And what this just shows is the tremendous variability for the interpretation of this picture from a set of diagnoses indicated here in categories in class 1 that are essentially benign and have no risk of invasive melanoma, no suggested risk, or some pathologists actually called this the same lesion in the diagnostic terms here on the bottom, invasive melanoma, which is associated with significant risk to the patient as well as treatment consensus would be required. So this just shows for one example how much variability can be seen. But, you know, she didn't just do this study on one case. They studied a whole set of different skin biopsies, and they show that this discordance among pathologists in the interpretation of melanocytic neoplasms is quite common across the community. And this is just an example here I'm showing in the central figure of about 60 different cases. So each column in this plot represents one case.. And the color of the distribution across the column shows the distribution of different diagnoses given to a case. And you can see for the vast majority of cases, a significant proportion of pathologists are classifying cases anywhere from these Class 1 classifications that are low risk to Class 5. And it's really remarkable that for any given case, for example Case 24, or any case you look at, you know, a third of the patients or so might be called the low-risk category and then increasing. And then as you move up, proportions of pathologists classify these cases into increasing risk. So you would hope that in every column you'd get a single color showing that all pathologists give the same diagnosis. But in fact, you're seeing quite a distribution of colors. And overall in this study, about 187 pathologists interpreted these skin lesion biopsies with an overall discordance to a consensus rate of about 45%. And then they also looked— well, now we can see there's a lot of variability across pathologists. Well, how much variability is there within just a single pathologist? So for 118 pathologists read the same samples 8 months apart, and then after 8 months gave a different diagnosis about a third of the time. So that resulted in an overall discordance rate of about 33%. So this shows tremendous inter- and intra-observer variability in the interpretation of skin biopsies. And Joanne Elmore has also done a very similarly designed study on the interpretation of breast biopsies. And that overall data is indicated here. And in this study, again, they found relatively similar results, where they found surprisingly high levels of both inter- and intra-observer variability with an overall concordance rate of about 75% for the interpretation of breast biopsies. But for some of the most difficult diagnoses, like, for example, the diagnosis of atypia, inter-observer concordance was about 48% for diagnosis of atypia. And again, in this study in breast biopsies, they found that intra-observer concordance, which is how well does an observer agree with themselves when looking at the same image separated by 9 months, that they only gave the same diagnosis after 9 months about 80% of the time. And for these very difficult atypical lesions, they only gave the same diagnosis as they themselves had given 9 months earlier about half the time. So this again just shows far too great inter- and intra-observer So across pathologists as well as within pathologists, variability in the interpretation of these breast biopsies. And this is showing some of the actual raw primary data from the study. So here's a set of about 70 cases that were all called atypical by the consensus panel. And then when they sent these out to a large number of community pathologists, they saw significant variability where for any case, there's a proportion of pathologists calling these cases benign, which is the tan color, atypical, which is yellow, or actually breast carcinoma, either in situ or invasive, which is either orange for in situ cancer or red invasive cancer. So for each case, there's just far more variability than we would hope. So this is a big problem, and it actually has been a problem for many, many years. So we might ask, well, could this get better? And are there other solutions? And the truth is, when humans view these images, really there has not been tremendous improvement for many, many decades. But one area where there has been great improvement is the ability of computers to interpret microscopic— well, to interpret images in general. And that's what's indicated in this plot, is just the improvement in computers using artificial intelligence. And this is showing over 2011 timeframe to 2017, the improvement of the best computational systems in something called the ImageNet Large-Scale Computer Vision Challenge, which is a challenge where computers are tasked with trying to classify the identity of different images. These happen to not be pathology images that are used in this competition. These are more like real-world images, like does this image contain a car? Does this image contain a cat? These are tasks that generally are done pretty reproducibly by humans. And you can see here that for these types of tasks, humans have an accuracy rate of about 95%. So errors are only made 5% of the time, whereas computers were really terrible at this task about a decade ago. So in 2011, the best computational system in the world was achieving an error rate of over 25%. So they would get these just wrong more than a quarter of the time. And these are for things that humans are very, very good at and can agree on. But you can see a dramatic improvement in performance from about 2012 to 2013. And that's really when there was this huge advance in the field of computer vision, or training computers to try to see what's inside images. And this was driven by improvements in artificial intelligence, particularly a type of artificial intelligence called deep learning. And deep learning is really defined as the form of artificial intelligence that uses neural networks with many, many, many layers within the networks. And these are hidden layers where an input of an image is, and then it's processed through these deep convolutional neural nets. And then the output is a prediction, for example, of what's inside of an image. And using this new technology of deep learning from 2012, which was really just developed between 2011 and 2012, caused this dramatic improvement of performance. For example, the error rate went down significantly from 2012 to 2013. And then incremental improvements in deep convolutional neural nets in general continued to get better year over year. At which point, between 2015 and 2016, the best deep learning systems were performing at a level that was as good, if not better, than humans at the interpretation of images. And what's really powerful about machine learning-based approaches is they can be very, very reproducible, where the same input essentially always will produce the exact same output for a system, which is not what we see with human interpretation of images. So I was part of a team of— a team from actually Harvard Medical School 5 years ago that took part in a grand challenge on cancer metastasis detection, which is a medical task within pathology. And the task in this competition was to see Could a computer automatically identify whether cancer was present in lymph node biopsies? And in this competition, there were over 200 teams. And the basic task was to take an image like this one here up at the top and to, to ask, is cancer present, yes or no, inside of these images? And we used a system called a deep convolutional neural net to try to build a system to make this classification. And this would be like an input image that the computer would get. And the computer is asked, do you see cancer anywhere? And you can imagine this is a very difficult task for both a computer or it would be for a pathologist. And what the computer can do is to look at every single region and even at a pixel level within the image and then to classify, does that pixel contain cancer or not? And if the computer thinks it sees cancer, it highlights that region in red, and if not, in blue. And you can see some of these hotspots that are very red within the image. If you look at a high power, you can actually see that this does contain a small rim of malignant cancer cells in a sea of many normal cells. So that suggests these systems can actually accurately identify the presence of cancer. And here's another very small hotspot that actually just contains something like a dozen of metastatic cancer cells. So this is a very powerful result and potentially a very useful new technology for pathologists. And then the best deep learning system in this competition actually achieved an error rate of less than 1%, which was significantly better than the error rates for pathologists viewing the same set of images. And this was really powerful data to show how useful artificial intelligence, and particularly the deep learning type of artificial intelligence, could be for the interpretation of microscopic images for routine tasks, like very important and routine tasks, like trying to identify whether cancer is present or not. And for a patient, the next really important decision, if the diagnosis of cancer is made, is what treatment was most likely to be effective. And one of the biggest areas of advance in the treatment of cancer broadly in many different cancer types has been the advance in immuno-oncology. And this was a study from about a decade ago now from Susan Topalian from Johns Hopkins and colleagues showing the safety, efficacy, and immune correlates of anti-PD-1 antibody therapy in cancer. And here's the particular drug that they were studying in this study called nivolumab, showing significant reduction in a lung cancer patient, both based on treatment with this anti-PD-1 therapy. But only a minority of patients in the study actually responded to the therapy. And the patients who did respond tended to have this protein called PD-L1 expressed on their tumors. And patients who are positive for PD-L1 showed a response rate of about 9 over 25 patients showed actual response versus patients who didn't express the protein in this study, none of them responded to therapy. So this looks like an important predictive marker for response. But the problem is it's very difficult to score reliably and reproducibly on images. And I'll show you why in a second. Particularly, it's very difficult to score on immune cells, where pathologists show moderate to poor inter- and intra-observer reproducibility for scoring this marker, as indicated in this study by David Rimm and colleagues that showed an agreement of only 0.19, which is not very strong agreement for the scoring of this marker. And the reason it's so difficult is because what we're asking pathologists to do is not only to identify whether cancer is present, but to actually try to estimate on a whole population of cells what proportion of cells are expressing this drug target. And the cells that are expressing the drug target are brown, and the cells that aren't expressing are not brown. And you can see there'd be a very, very difficult task for anyone to look at a whole set of thousands of cells and try to estimate the correct proportion that are brown, yes or no. That's the type of task that you can ask a well-trained machine learning system to do, and they can do it very, very reliably. And actually classify every single cell into both its cell type. Is it a macrophage, which is a type of immune cell? Is it a tumor cell, which is the cancer cells themselves? Or is it a lymphocyte, which is another type of immune cell? And the AI system can both accurately classify the cell type and then classify whether the cell is expressing this protein, which is brown, and then that is indicated by this yellow border, and then can just summarize exhaustively over all the cells and compute the exact percentage that are positive. So this can make this interpretation of this important biomarker more accurate and more reproducible and to help oncologists and pathologists better make the diagnosis, which we hope will lead to the best therapy that will be selected for patients. And this is one study that we published, presented at the American Association for Cancer Research in 2020. And it now is in a manuscript that's under review for publication showing the potential power of this technology for patients.. And the key groups to look at, all of these patients received anti-PD-1 therapy. The patients who respond the best in the blue lines in both melanoma and urothelial, which is bladder cancer, on the top, the patients in the light blue category showed the best survival. The patients who are negative for the protein as measured by both machine as well as by pathologists here in purple, they showed the worst overall survival. So they had the most aggressive disease. And these patients in orange were patients who were called differently between machine learning and human.. And this is a larger number of patients were called positive, potentially because the system can exhaustively score every single cell. And these patients actually showed overall survival that was more similar to the truly positive patients, suggesting that these are patients that, that actually do benefit from this therapy. And there were patients that were only identified in this study through the machine learning-based analysis. So this is just one example, but broadly, we have hope that AI-powered pathology will really advance the field of precision medicine. By providing pathologists with tools to enable them to provide more accurate diagnoses. When deploying these systems within clinical trials, the use of this technology, we hope, will lead to more life-saving drug approvals. And then the deployment of these systems broadly in the world at large will enable doctors to more effectively match patients with the best therapy for their disease. And broadly, we really hope every patient will have access to the right diagnosis. So that they can then receive the right therapy. And that's my last slide for this talk. So I'd like to pass it over to Dr. Viswanathan, who will lead the next question and answer session. And thank you very much for your attention.

Hello, everyone. That was an outstanding presentation. And I want to thank you all for— and especially to Dr. Beck for his great discussion. I'm going to begin with a few questions and ask some of these questions to Dr. Beck. So incorporating some of the questions from the audience, there was a question about how were these cases chosen? Were they random cases, or were the cases chosen to be difficult?

No, that's a great question. So particularly for those two studies I showed earlier, they were a mix, but they were they definitely were overrepresented with difficult cases. So it was a distribution of cases that were normal, cases that were invasive cancer, which there tends to be the greatest agreement between, but they were overrepresented for these sort of borderline cases in the middle that were either atypical. So definitely the disagreement rate seen in that study may be higher than is seen in practice. But it still did cover the full, the full spectrum and was run sort of a study by cancer epidemiologists. So, so it does represent the different error rates within these different areas. But definitely more difficult categories were overrepresented in this study than what you would see if you just took samples at large. But this, this does sort of allow us to understand the error rates in in all categories—normal, cancer, as well as this sort of in-between categories. And the in-between categories were overrepresented.

Excellent. So the, you know, the importance of what the learning set is, right? What the information input is, is always paramount to doing these kind of tests. So another question that was asked from the audience is, how often do pathologists disagree with the prediction of the neural net, right? If you have individual pathologists evaluating the results of these tests, how often do they overrule the machine prediction versus changing their diagnoses?

Yeah, I know it's a great question. It's very contextual, so there's not going to be an overall answer, but I can give you some sort of general themes which I think support a lot of sort of research in areas like behavioral economics or the idea of nudging people can very much impact what they end up saying. And, you know, obviously we hope in a positive way. So studies we've done, for example, where we've looked at this very extensively in other difficult areas like the interpretation of liver biopsies, where the disagreement rate might be something like a third or 33%, which is very similar to also what we saw in this heading of breast biopsies or skin biopsies in those very elegant studies by Joanne Elmore and colleagues. So if the background disagreement rate between people is about a third, or even people versus themselves— some of the very striking data we saw was that pathologists even disagree with themselves when they look at the same images 9 months apart. If that background rate is something like a third, if they're actually provided with the guidance of a machine learning system that's been very well validated and actually shows, as I showed in the images, exactly why it made the predictions— and it does that by exhaustively classifying every single cell In the work we're doing today in liver disease, for example, we see that the disagreement between pathologists and the computer is something like less than 5% of the cases. So kind of like nudging people to a diagnosis and then providing the evidence for why that's happening, often the disagreement is less than 5%. And then the question is, if you do an adjudication panel, how often does that disagree? So another set And then in some of those cases, the actual next viewer agrees with the machine. In other cases, they'll actually agree that, yes, I agree with the human. And that's a subset of the 5%. But there's a dramatic reduction in variability by supporting pathologists with this. So I would just say in certain cases we've seen the disagreement rate goes from 33% to less than 5%, which is a really, really big drop in disagreement by having machine learning support the pathologist.

There is so much individual variation. You know, I'm in cancer, and in cancer pathology, things like grade or Gleason score for prostate cancer are highly, highly variable. So I can see how, you know, there's an element of this that is about standardization and then another element that's about interpretability. And that's really what you're getting at, is kind of trying to get that standardization component. And then there still has to be the, ultimately, the physician, the pathologist clarifying and agreeing and saying, you know, we're not going to overrule the machine prediction. So the next question is really quite interesting. It's about how is the AI pathology field addressing regulatory changes such as FDA approval for these tools?

Right. Yeah. So the FDA plays a really critical role in in the approval of both therapies but also diagnostics. So we're working very closely with the FDA, um, on the approval. And the FDA has a history of approving image analysis algorithms, so there's actually dozens of algorithms for very specific cases that have already been approved historically. So there's a great process for how they approve it. At a high level, there's really two things. They first want to show it's analytically validated, so it's measuring what you expect it to measure. So there's a set of cases that there's a ground truth diagnosis. Often that's done through consensus of what pathologists think. And then you have to show that your system agrees with that with very low error rate, and an error rate that typically is lower than what a pathologist would get by themselves. And then there's clinical validation, which is validating how this would actually be used in clinical practice. And again, if it's an accuracy test, you have to show that, that during clinical validation that the test achieves a certain success criteria. And then after that, the FDA can, can say that this is providing enough evidence that it can be safe and effectively used. And there was just recently actually an approval of a deep learning-based system for the interpretation of prostate biopsies. And we expect there'll be more and more in the coming years. So there's a pretty good regulatory framework for the approval of these sorts of tests.

This is— I'm going to ask because I'm just curious how How does one allow there to be significant heterogeneity within a dataset to account for all of the diversity of a population? You know, in a global community, there may actually be significant differences, even though something may be called, say, you know, a cancer of the skin like a melanoma. Is that different in a population that's of African origin, a population that's of Icelandic origin, a population that is heterogeneous in the United States, and how does something like an AI-based training tool account for that significant heterogeneity by gender, by other cultural differences, et cetera?

Yeah, I mean, it's a really important question. It really is not one that's specific to AI. It's one that is just broadly important that the validation dataset is well represented of the population. Where the tool will be deployed and that physicians are well aware of the population that it's been validated in and really incorporate that into things like the approval and the label of the new diagnostic. So just as you mentioned, it's really important that those datasets are well-represented in training so that we can have a high confidence that the system will generalize well across populations. And even more important, I would say it's important that the populations are well-represented in the validation dataset So we really have good insight in how robust these systems will be across diverse populations that you can expect to see in the real world. And then third, it's important that physicians are educated about, you know, what has this actually— which populations have these systems actually been approved in and to not misuse systems in populations that we don't yet know what the performance characteristics really will be. So it's a question of training the systems appropriately. Validating them and providing that data to the FDA and other regulatory agencies for the approval, and then educating physicians so that they do use these tools correctly in the real world.

Another set of questions that are coming through are kind of linked, which is really— it talks about the slow adoption in pathology compared to, say, radiology, which I think is also, you know, a progress and evolution. But the question is, is Is the slow evolution and adoption in pathology due to lack of data standards, preference by pathologists for microscopic images?

What's your theory? That's a great question. So one is the technology has only recently actually gotten to where it can be really, really good. And we're just now seeing these things start to be FDA approved. So I actually think there will be fast, relatively fast adoption over the next 5 years or so because If you look at the core advance was in about 2012, where deep learning version of artificial intelligence outside of medicine really showed advance. Geoff Hinton and others were some of the leading researchers about less than a decade ago now who made the first advances, I would say. And then I was actually part of the team in 2016 or so that really on the research side show how powerful this could be for pathology. And then over the past 5 years, we and others have been working on really validating these, making these systems very robust, And then beginning to put the experiments together and provide the data back to the FDA, which is really happening now, to get things like FDA approvals and other types of approvals where some of these could be used as laboratory-developed tests, which have a different approval. And it's going to really be over the next 5 years where first we have the regulatory approvals and the validations in laboratories that can work and solving for the clinical workflow. And certain labs will start to adopt them. And then the other is the reimbursement and the payment to, you know, which also can drive physician behavior to put those in place where tests that are providing medical value are providing, you know, physicians have the ability to actually pay for these technologies and use them. So I think it's a combination of the regulatory factors, actually having just the technology available, clinical workflow, and then sort of the reimbursement models, which also drive clinical behavior. Um, I, so I think over the next 5 years we're going to see all advances in each of these areas, uh, that will really change the practice of medicine. Um, but that's why, you know, changing behavior takes time. Um, but we're very much seeing, um, uptake, you know, I think we will be seeing increasing uptake over the next 5 years. Uh, and I think the biggest factor is just, it's The technology is quite new, and I actually think it's less cultural and more that they're only now really being developed and rolled out. A lot of the research that I showed was still quite early and not fully developed clinical products that can impact clinical care.

This is a technical question that someone asked, which is, as you're looking for specific feature extraction in the early layers of the network, do they look more generic, or was this a transfer learning paradigm?

So yeah, a lot of the— it's not transfer learning, the stuff that we're doing, because typically we have access to very large datasets and are able to build very large datasets, so we don't necessarily need to do transfer learning. And just for others in the audience may not know, transfer learning would be, for example, you train a a generic, in a way, deep learning system that can identify cats and dogs and hot dogs and colors, things that are not specifically related to pathology but just general visual tasks. And then you could start with that model as input and kind of fine-tune it and refine it on pathology-specific tasks. And in certain cases, when you have limited training datasets, those systems perform quite well. What we've shown is if the training set is large enough, you really don't in our experience, get advantages from transfer learning if you have a large and robust enough training set. So we haven't really been, at least at PathAI and in our work, been using transfer learning from other settings. And then in terms of what features are being picked up in the network, that's very interesting. And it's kind of— and often the systems do actually focus on the same things pathologists have been focusing on. If it's trying to do a task that pathologists are already doing, it often focuses on the same sorts of features in the image the pathologists are using. And one thing we've learned in our work is pathologists have learned a ton about what actually is clinically important over the past 150 years. So we think there's actually incredibly strong knowledge that physicians have. And then the networks, kind of because they're learning from pathologists, they learn to focus on the same things. But for some new tasks, like the thing I was showing at the end about predicting drug response, which are very new, and some of these drugs have only been around for, you know, a couple of years, Often these systems actually clue in on things that pathologists have not previously been thinking about. So there's kind of— those are the two big buckets of systems, either reproducing what we already know is important but making it more accurate and reproducible, but also we can really use these deep learning systems for discovery and actually discovering new components of the image that weren't previously thought to be important.

Another question here. It is well understood that there is a significant difference of opinion about melanocytic lesions among pathologists. I understand AI will reproducibly give the same result. How do we know if it's the correct diagnosis? Great question.

The, the only way we know is through clinical studies and outcome. And for example, like I showed at the end, the overall survival being really the true metric. So being reproducible is very important, though, in its own right. So I do think reproducibility and standardization is just— is really important. By itself, just because that becomes the basis for how you're going to conduct future studies. And also, you know, patients should have access to the best diagnosis in a reproducible fashion, and best being what the experts think, which is what we use for the ground truth when we don't have survival. But at the end of the day, you know, patient survival— and just to be very clear what that would mean and how you would study that, you would have a population that you study population who got their treatment done without the use of AI and follow how well they do. And then you have a separate population where their care was augmented by AI, and then you evaluate whether that second population actually does better in terms of survival compared to the first population. And if they do, you can infer that this system is actually better. It's not just difference of opinion. It is better against an objective ground truth and things like that. Overall survival are a very strong objective ground truth to show you're really making a positive impact on patient outcomes.

Okay. A question near and dear to our moderator, Dr. DeVivo's heart, which is how can epidemiologists get involved in this field?

Great question. So it's a wonderful field for epidemiologists. So epidemiology really relies on things like pathology for classifying patients into categories. So I would say partner with folks in your institution or others who are using sort of AI-powered pathology to make these classifications more accurate, more reproducible, and even more data-rich. And then for patient samples or studies that are being done on diseases like cancer where tissue pathology plays an important role, I do think it's partnering with folks using AI-powered pathology to incorporate that into study design for the way that you're going to classify the diagnosis, not just drawing only on pathologists, but pathologists augmented by AI. And then AI can make not only these sort of binary existing categories, but can create entirely new categories by measuring samples in new ways.

Another kind of technical question, which is, how do you clean your data?

Oh yeah, so the dirty truth in machine learning is cleaning the data is probably 90% of the work. There's very wonderful frameworks available now for labeling data and for making the machine learning models. Companies like Facebook and Google have made a lot of open source tools for building models, but what they haven't done is for cleaning data. There's a lot of automated cleaning and then manual cleaning that we do. That includes doing things like building machine learning models for automatically finding artifacts for automatically finding blurry regions of images, for standardizing images so that it doesn't matter which scanner the image was generated on, that we have a way to use machine learning to analyze images that are produced by different scanners. And then sometimes there's manual curation. So people making sure the metadata makes sense, which is sort of the labels on the data. So there's a whole different set of steps of data cleaning that become very important for training these systems effectively and then for deploying them effectively. Effectively. And we try to use machine learning wherever we can for building models to automatically clean the data.

I kind of focused your specific question— immunohistochemistry can have technical issues, especially when, say, a bone biopsy goes through decalcification. Is AI addressing this sort of issue?

Yes, it's almost a good follow-on to the last question. It's— I think it's a wonderful use of AI. Is to remove artifacts automatically. And I didn't fully answer it. The way you actually do it is you train systems. So you have humans, for the most part, identify this is artifact, this is not artifact, this is artifact, this is not artifact. And then, you know, at PathAI, we have over 12 million of these types of annotations. And then you can automate the process of identifying things like decalcification, auto— artifact due to different tissue processing methods. AI is very, very good at those sorts of very tedious and repetitive tasks, such as identifying artifact, removing artifact, and making sure that you're focusing on the in-focus salient part of the image. There's a lot of preprocessing, much of which is done by machine learning, that goes into then ensuring that the model is then really paying attention to the most relevant parts of the image.

This is again somewhat related as a technical question, which is about, do you see the need to do any normalization of the image data across data acquisition systems? Yes, absolutely.

It's a really critical part. I would say going from the world of academia where I was in my world from like say 2011 to 2016, and then the past 5 years, the world of industry. When you get into industry, it's just so much more important and I think more well done and there's enough resources to really do things like normalization such that you can then deploy in the real world where you encounter tons of artifact and variability. So it's incredibly important, and I think something that's often not done enough. And I think the biggest cause of lack of reproducibility of a lot of research is people not carefully doing things like normalization and not having robust enough training sets and robust enough validation datasets, because normalization is really important to ensure robust deployment of models. And then the, you know, the thing that's super important too is to actually do this sort of validation of models where you have— where you can actually test whether the scanner will produce enough noise to mess up the performance.

So yeah, it's absolutely critical. How many patients were included in your study? How many images do you get per patient, how many predictions are you doing per image?

Well, we've done lots— I mean, lots and lots of different studies, but the one that we showed here, um, so the, the data that we showed on validation was on the order of hundreds of patients who weren't used in training, where we actually had overall survival data. So we showed in the one plot I showed here was bladder cancer and melanoma on the order of hundreds of patients in both studies across In this case, it was all the same treatment. It was from a controlled clinical trial. And then what was used in training— again, on the order of hundreds of samples were used in training. But the important— and in general, in clinical trials, there's one slide per patient that is used in this— in the analysis. But the— these pathology images are so large that they often have literally hundreds of thousands of cells per image. And then on the training side, we used in this study about a quarter million annotations on these images. So although we might only have hundreds of images, we can get essentially hundreds or so of annotations per image. And we can end up on the scale of hundreds of thousands of annotations. And what an annotation is, is a pathologist circling saying this is an immunohistochemistry positive cell. This is a negative cell. This is a cancer cell. Immune cell. So they're hand labeling images. So I would say typically on the order of hundreds of slides or thousands of slides, hundreds of thousands of annotations. And then on validation, it's typically something like hundreds of patients.

And so somewhat related, but is it reliable that you can differentiate benign from malignant? Using deep learning? Because obviously, you know, that distinction is so important. There's a medical legal issue if you classify something as benign, but it were truly malignant and were found on a re-review by some expert at some point to be called malignant, you know, and that those critical junctures, you know, how reliable is an AI tool at classifying benign versus malignant? Yeah.

I mean, the answer tends to be, you know, kind of it depends. And it depends on how the data is going to be used. So I would say a lot of the work we do is exploratory research. And in which case, we often do the analysis in a fully automated manner. And how reliable? I mean, there's both analysis at the patient level and then at the cellular level. And we typically try to build models that perform even better than a human, and even humans aren't perfect. So for example, for classifying cancer cell versus lymphocyte, you know, we can be close to, you know, 99% accurate, for example. So the error rate for an individual cell might be less than 1%, and when you're summarizing over a population of cells, the error rate at the slide level would be very, very low. For certain patients. And then there's that whole area of ambiguous, not totally normal, not completely cancer. And then the accuracy rate or the agreement rate might be more like— the agreement of an expert with the system might be more like 75%. So it really depends, I would say. And there's no totally general answer. And it's always important to validate on new data. But in general, we aim to be even better than a human.

And there's always going to be errors no matter what. Someone did ask for another technical question, which is, can you elaborate on how you developed this and how did you decide on the architecture that was used for the deep neural network?

Yeah, sure. So there's lots of really good high-performing systems now that are available, and we, you know, in our, in our company have, you know, we have a team of dozens of machine learning scientists who evaluate many different architectures. And there's usually a trade-off between kind of complexity and how long a system will take to train and how difficult it will be to deploy, and then performance. And really, we tend to just optimize for performance. And the way we make model architecture decisions is performance on unseen data is kind of the single rule in all of machine learning is how well do you perform on systems that you've never seen before? And then what's the cost in terms of training and deployment time? So those are some of the trade-offs we think about when picking model architectures. And I would say for other— for folks just joining the field, that we're living in a really awesome time because there's just so many publicly available architectures now to begin with. And then to really decide which one is best for your particular task based on model performance.

How often do the different deep learning models disagree with each other, and do we test models against each other?

How often do they disagree with each other? Not— I mean, in general, I would say not very often. The differences in these models that we work on, at least model architectures, tend to not be great. But I would just say, so I would say in our system, because we control everything so well, and there's so many different systems, and if we're using the exact same training dataset, performance tends to be pretty consistent. If the only variable that's being changed is the architecture. But in general, the disagreement between models will be as great as the disagreement between anyone. If you're talking about different systems trained on different sets of samples or different scanners or some people didn't normalize, others did. So in general, just knowing that something is machine learning based or AI based is not helpful at all. In the same way that you can't tell how good a person, a human being is going to be based on I sent this to a human and they told me it was cancer. I sent this to a human— are they a pathologist? Have they been trained? How well trained? What are their credentials? Are some of the things we use for humans. Same thing is true for models. Is this being produced by a very credible group? Is there clear documentation of the training dataset, the validation dataset? Has this been approved by a regulatory body or has this been developed under design control? Is this being compliant with things like CAP and CLIA credentials? So there's all these different variables. I would just say machine learning models disagree with each other as much as humans disagree with each other. But to the first question, for us, because we're controlling for so many factors, sometimes just varying a model architecture will not have a big impact on performance. So it's very long-winded, but I would say— The devil's in the details, and certainly machine learning models can disagree all the time, and you really have to go deeper and say, you know, how well was this trained, how well was this validated.

Exactly. I think that's what we're learning is that, like with anything, you know, it's— everyone gets the basic concept and loves the idea, but it's all about how it's done. And really, you know, you've dedicated your whole career and all of your effort, and, you know, there are many, many people interested in entering this field in order to make it more robust and perfect the science, which is really what this is all about, so that it's standardized. Similar to that, there's a comment that it says, quote, deep neural networks are known to be very sensitive to small superficial changes in the data. So for example, changes in illumination, resolution, stain coloring in the case of histopathology images. Human doctors are presumably good at filtering out these nuances. So, you know, what's the state of the field in computational histopathology for dealing with this lack of robustness on deep neural networks?

Yeah, I would say it's a great question. If things aren't done robustly or well, they can be— I mean, I would disagree somewhat. I think deep learning is far better than other tools in general in terms of dealing with artifacts. I mean, that's why, like, when I showed the data early on about the performance in that system, the Chameleon Challenge, deep learning systems crushed every non-deep learning system, largely because they can deal very well with artifacts. Or if you look at what happened with the ImageNet Large-Scale Computer Vision Challenge, why it did so much better from 2012 to 2013 and got even better from 2013 to the present, it is because deep learning actually is very good at dealing with artifacts when it's done with expert practitioners. And so I would say, you know, these systems are actually some of the best systems. But I totally agree that in clinical practice, I think we absolutely want and have to have pathologists making the final call because you never know, maybe there's an artifact that the system hasn't seen before. And humans are very, very good at dealing with artifacts and weird things, dealing with things they've never seen before. So I think there's very much complementarity between the types of mistakes humans make versus the types of mistakes machines make. And humans tend to be very good at sort of judgment and categorical things, whereas computers are very good at exhaustive quantification. So there's awesome complementarity between the strengths and weaknesses of both. And I think for clinical decision-making, absolutely, at the end of the day, you want a human seeing the output of the machine, seeing what the raw data looks like, and just making the final call. And that's ultimately going to be the safest and the most effective. For patients. On the research side, sometimes just the pure scalability of machine learning will trump it, and it's just not going to be worth it to have a physician re-reviewing everything because there's just strength in numbers. So I think, you know, pure research, a lot of it machine learning can do all the work, but for clinical care, I think you're going to want a human making the final interpretation, being assisted by machine learning.

One last very quick question: what percent of pathology is currently using AI in the clinic? And if there are barriers to why it might be lower than you'd want, how are you going to fix those?

Sure. So the truth is today, very, very little. So much of this is still research. 99% is still not used by AI. I actually think we look ahead 10 years, it's going to be the opposite. 99% will be AI-assisted. What's the biggest barrier? I think the biggest barrier is the technology is so new. As we've heard about from all these questions, it's really hard to do this well. It's only very recently that we've had the ability to do it well. And we're only just now seeing this technology really cross the line where it can do it, where it's really making doctors better. But the awesome thing about machine learning compared to, I would say, what we've seen with humans is once it starts getting better, it's going to keep getting better. And more and more data is going to keep making it better. That's why I actually am quite confident we look ahead a decade, we're going to say 99% of things in the US and in research-rich resource— where resources are available, you know, they're going to be leveraging this. And then I think the big challenge— and 99% of people will be using it— and then the challenge will be how do we distribute this and really make it ubiquitous and democratized around the world, you know. And that's probably the next big challenge. All right, thank you so much, Dr. Beck. Thank you, it's been a pleasure to be here and to to be able to take part in this seminar.

Yep, outstanding answers to all these really complex questions from the audience. Thank you also to the audience for providing a really great distribution of questions for us to talk about today.

And I see Dr. DeVivo is back on. Thank you. If I may, I'd like to refer to friends and colleagues. Thank you, Andy. Thank you, Akila. This concludes our program today. I want to thank you both, Andy, for a thoughtful presentation and perspective, and Akilah for serving as our moderator. You were terrific. And the audience, truly terrific questions. And clearly there's not enough time to really answer all the thoughtful questions that were asked. We hope that you will join us for the Ethel and David Jackson Next in Science program on climate change on January 31st. And we hope I hope you will join us for the next science lecture series program on artificial intelligence with Dr. Timnit Gebru. Should be a terrific talk. That program will be held on Tuesday, March 8, 2022. Today's program has been recorded and will be posted on the Radcliffe website in about a week or so. For information on upcoming Radcliffe virtual programs and to see videos of past events, please visit www.radcliffe.harvard.edu. And I want to thank you all again for joining us, and take care.