[00:47:00] Gil Elbaz: Okay guys, I have here someone very special with me today. Idan Bassuk, the VP A.I. Of Aidoc. Idan started four and a half years ago as an AI Algorithms and Software Engineer. Today, he leads the AI group, a 90 person group, which concentrates all of the efforts required for AI, from dataset development, through algorithmic research and up to deployment and continuous monitoring in production, at a scale of over 500 medical centers in the world.
[00:01:14] Before joining Aidoc as its first employee, Idan served for 10 years in the Israeli Defense Force. Idan started in the elite technological Talpiot course, which he graduated at the top of his cohort. Later, Idan served as a team leader in a special operations unit.
[00:01:31] He then finished his service as the head of a technological section, leading the defensive technological projects, which were awarded Israel’s most prestigious defense award for the success in detection of terror tunnels, crossing the border into Israel.
[00:01:46] Idan, it’s a pleasure to be with you.
[00:01:48] Idan Bassuk: Pleasure to be with you. It’s really exciting to be here today. Nice meeting you.
[00:01:52] Gil Elbaz: So going from tunnels to medical, that seems like a huge jump in different fields. How did you make that jump? What was your path into the medical domain?
[00:02:03] Idan Bassuk: Excellent question. So I think, uh, first of all, maybe it’s more similar than you would have imagined because eventually finding tunnels underground, it’s essentially imaging of the underground, medium, finding something behind something that the human eye can’t see in the medium into which the human eye can’t see similarly to imaging of the human body.
[00:02:26] Gil Elbaz: I guess it’s also extremely hard for a person to infer small signals that you would get in that kind of data.
[00:02:34] Idan Bassuk: Of course, it’s extremely challenging. And that’s why took several decades to find solutions, to find the first solutions these problems. But eventually in, I think that it also connects to what we’re going to talk about here today.
[00:02:48] What enabled us to solve it. Wasn’t only technological progress and in many, in many aspects, most of the technological building blocks that we eventually used for the detection of tunnels existed several years before we succeeded in developing these systems. And many of these things were working with the correct production, great methodologies, and really tackling the problem from a holistic manner.
[00:03:15] Not only focusing on the algorithms or on the sensors, but finding a holistic organizational and operational solution that combined in fits everything together. And I think. I took many of these insights with me to Aidoc, not only in the algorithmic imaging aspects, but also in how to tackle the problem of bringing AI to production in the real world at scale, by tackling it from a holistic direction and, uh, using engineering and production grade concepts to make it work
[00:03:50] Gil Elbaz: interesting.
[00:03:50] And I’m a huge fan of Aidoc, but for our listeners, I’d love for you to tell a little bit
[00:03:55] Idan Bassuk: Mutual about Datagen, of course.
[00:03:57] Gil Elbaz: Thank you. Thank you very much. I’d love for you to give our listeners a bit of insight. What Aidoc does, what kind of value it delivers? Who does it focus on?
[00:04:07] Idan Bassuk: Sure. So Aidoc is the leading provider in AI for medical imaging. Medical imaging, like CT scans., X-ray, MRIs, are scans, interpreted by doctors that are called radiologists. This is the specialty of doctors that specializing in reading and interpreting and diagnosing based on these medical images. And it’s a really, really complex field. Our goal at Aidoc, we have actually built system around the concept that we call it being always on.
[00:04:40] It means that we are always on running in the background. We don’t wait for the radiologist to send us a question about a scan, but we are listening to the databases of the hospitals. In many cases, we are even connected directly to the scanners themselves. And once this new scan gets acquired in a CT scanner, for example, we automatically identify it, analyze it with AI algorithms to detect different types of medical conditions, such as brain hemorrhages, such as fractures the spine, such as strokes in the brain, are really critical and life risking medical conditions. And we often do this analysis much before the radiologist even opens the scan. Sometimes hours and even a day before the radiologist would even open the scan otherwise to analyze it. And our goal, we succeeded in doing that, is to drive the radiologists to get to the most important and the most critical patients earlier.
[00:05:41] And the second thing is that we actually caused improved patient outcomes, both by that. And by proving the radiologist quality, making sure that even when they are tired and even when the radiologist that is, has not specialized enough yet. Or just like came, uh, with a bad mood today from, uh, from home. We help them. We’ll make sure see the findings and even the smallest findings. And we use actually the advantage of AI, It’s never tired and it can always concentrate on each and every pixel of the image at the same level of concentration.
[00:06:17] Gil Elbaz: if I try to encapsulate that you guys have a super tool for radiologists, that is always on always active, always looking for problems in these different scans, coming from a ton of different devices.
[00:06:33] And when it sees a problem, it raises a hand and it says, Hey, look at me. And also when they’re actually analyzing the scans themselves, the images themselves, it helps them in aids them throughout the analysis process to make sure that they’re always spot on.
[00:06:48] Idan Bassuk: Yeah. Yeah.
[00:06:49] Gil Elbaz: That’s amazing. It sounds like you guys save a lot of lives. Do you have any data on that?
[00:06:53] Idan Bassuk: Yeah, sure. many of the medical centers that are using our products are evaluating the performance of our AI on their own data before they acquire the product before they purchase the product, because they want to see and be convinced that it really works on their own data.
[00:07:14] Even though that we are already installed and running, in the really the top institutions in the world, like, Mayo clinic, which is ranked almost consistently every year, it’s the number one hospital in the US and then Mount Sinai, which is also one of the top 10 hospitals in the US By the way, I think several dozens of percent of the scans in the world are actually analyzed by radiologist.
[00:07:37] that don’t work for hospitals. They work for private groups, private companies that give service to hospitals and we are
[00:07:45] Gil Elbaz: Interesting.
[00:07:46] Idan Bassuk: Yeah. And we are also pretty successful in that area as well. And we have sold and we are working together with the top groups in the world, including the number one largest group in the world, Radiology Partners, which we signed a strategic contract with, which just to give the sense they have around 3000 radiologists in their company.
[00:08:08] So we have already proven that our product gives value, but you’ve asked if we’ve seen it, and have we quantified our value? many of the medical centers that we serve, even though we have these set records still want to see are as performing on their own data.
[00:08:26] So we do retrospective studies. On years of data months or years of data, depending on what needs to be done to be statistically significant. And we’ve done it on millions of scans already, even as part of researches, which are conducted independently by these medical centers, they published dozens of academic papers on it.
[00:08:47] And we have seen in these controlled researches, even, that we are improving the patient outcomes by reducing the missed detection rate by radiologists. And we have a very good rate of, or a ratio between our sensitivity or what is known as the recall and the number of false positives that we provide.
[00:09:09] Gil Elbaz: What would you say is main challenge now for Aidoc? Because it seems like a no brainer. If I was a hospital, it makes sense. It just, it would save them money. It would save time. It would save lives. But what is the main challenge for Aidoc in the next two years forward?
[00:09:23] Idan Bassuk: It’s a great question because with everything I told you about so far and you know, the best hospitals and the groups in the world that are already using and loving our product on a daily basis, by the way, one of the great things is really on a daily basis, we get emails or WhatsApp messages from doctors that are actually using our products or heads of radiology departments, et cetera, that are giving us examples of patients, which, which lives we saved or helped save. Uh, which really is an amazing feeling.
[00:09:55] And it might sound like everything is behind us. Like we have, you know, like in a executable file that you can now just sell it to 5,000 hospitals and, and get it over with. And it’s really not, not that.
[00:10:11] The challenge in getting from zero to one hospital and the challenging getting from the first 10 hospitals and the challenging getting from 500 to 10,000 hospitals, there are deep technological challenges in being able to maintain the same level.
[00:10:28] We think we have proven that we have product that gives value that gives quality and that the customers and doctors want and hospitals want to use, but doing the same thing and giving the same level of providing the same level of quality at a scale of 10,000 hospitals, which is our next target, uh, requires a lot of technological work from many different directions.
[00:10:50] one of the major challenges in AI for medical imaging is the data variability. There is no standardization in how to scan patients, even if you know that you are looking for a specific medical condition, which often you don’t know, like a brain hemorrhage, different medical centers will configure the scanner’s physics, even things like the x-ray frequency, differently between different medical centers and the images look really differently between medical centers. As you scale, you need to be able to give the same quality on many, many different appearances of these scans.
[00:11:29] Gil Elbaz: So similar to autonomous vehicles where there’s a very long tail of uncertainty as you grow and look at a wider target visual domain. Is there a similar long tail in the medical space?
[00:11:42] Idan Bassuk: definitely. In these terms of data variability, in each type of scan like CT x-ray MRI. There are dozens of parameters in each of these types of scans that can affect the image appearance. And the number of appearances is not dozens.
[00:12:00] It’s millions because it’s the Cartesian multiply of all of these different parameters. Each parameter can be configured independently of the others, and eventually you can get a very different appearances. Sometimes The differences will be visible and significant to the human eye.
[00:12:17] Like the image can be much more noisy. The contrast can be greater, but sometimes the differences are very subtle and not clear to the human eye, but as we know, well, while deep learning is, uh, probably the most robust type of algorithm developed today, it can still be very sensitive to surface anomalies.
[00:12:40] And surface irregularities, even those that are invisible to the human eye that are different from what it was trained on and tested on. And that’s why it’s very important to, keep adapting your algorithm, to all the types of variability that you see in the real world.
[00:12:55] Gil Elbaz: I think that one of the nice things and interesting takeaways that I see from this is that you guys see a certain situation. You guys see that there’s a ton of variance, there are so many different devices and different methods of, of configuring these devices, which obviously creates a very challenging visual domain.
[00:13:12] And you guys don’t say, okay, let’s stop everything and create a device or tell the doctors how to work. You guys come to them and say, okay, we can solve this problem, even though there’s an enormous amount of variance.
[00:13:25] Is there in the future maybe ways to work together with the doctors to help them, help you, help them?
[00:13:32] Idan Bassuk: Today. I think the doctors are already helping us help them. In many ways. They are giving us feedback on our product and helping us navigate it and improve it in the ways that are more, most meaningful to them.
[00:13:46] And this is probably the most valuable thing that the startup can have. We are installed in over 500 hospitals, which really love our product and love to give feedback and help us improve, including the top medical centers in the world. And by the way, I think at least already six medical centers in Israel, like Sheba and Ichilov, so,
[00:14:07] Gil Elbaz: so I’m personally happy to hear that.
[00:14:09] Idan Bassuk: Yeah. So it’s a really great feeling by the way, to know that you potentially, you’re going to save life someday of your family members and friends. So, yeah,
[00:14:18] Gil Elbaz: So deep learning is the most robust and algorithmic solution for these kinds of computer vision challenges.
[00:14:24] And we’ve seen it do amazing things in the last few years. And then the big question, and this always arises when people talk about medical data, which is around explainability. Is this important? Is there a certain level of explainability that’s necessary? How do you guys go about explaining, talking with the radiologist, about what the algorithm is doing behind the scenes?
[00:14:47] Idan Bassuk: Great question. I think that explainability in many cases is very important, but it’s also a very hard and not very solved problem compared to other, types of problems inside AI.
[00:15:00] I guess that’s why raising this type of question. And I think that’s for a startup and any company building a new product and a breakthrough product in general, it’s important to be as focused as possible and define the problem in a way which minimizes the number of unsolved problems that you need to solve.
[00:15:20] And I think that at Aidoc, one of our insights from the early days was that we build our product in the correct way, explainability is less of an issue for us in many of our use cases. Because once you detect a finding and you drive the radiologist to look at that finding, she can see for herself, and she can say, okay, this is a real finding. So she doesn’t need you to explain why you think this is a finding. You already convinced her it’s a finding, and she can move on, and take care of the patient.
[00:15:52] Gil Elbaz: I think that’s an amazing explanation. You’re saying we want to make the product in the best way possible. So that explainability isn’t a barrier.
[00:16:01] And I think that that’s probably the best way to go today because it, like you said, it’s an unsolved area and there’s a lot of work going into trying to make the findings of deep neural networks, more explainable, trying to make what the network is doing inside of the black box, more explainable.
[00:16:18] But today we see that, you know, deep features that have highly entangled representations of what’s happening in the real world are really driving a lot of the decisions. And there are a lot of people that hypothesize that it’s similar to the human brain, right? That we don’t have a specific neuron that tells us if it’s a finding or not.
[00:16:36] And so we see that these algorithms work and I love the approach of making sure that the customer or the radiologist that’s using the platform in the end feels confident and has the power to actually make the final decision.
[00:16:50] So you step into this new world of medical data, right, and you define certain computer vision tasks that you want to solve, be it detection, segmentation, classification.
[00:17:02] How do you go about understanding the state of the art today? Especially with the endless papers that are constantly rolling out from the academia into our arXiv feeds?
[00:17:13] Idan Bassuk: Great question. So I can say a few, a few things about that. I started in the field of deep learning, uh, four and a half years ago. And I think that there are no shortcuts. You need to read a lot of papers and you need to dedicate time to read the papers and it’s important.
[00:17:28] And there, there is a lot of gain knowledge and a lot of intuition, both about the algorithms themselves and about what is the right way to go about and perform research in this area. And also reading more papers gives you the context to judge papers. You read more correctly because more often than not actually I think that you see like misleading comparisons between algorithms and thinking yeah.
[00:17:53] Cherry picking or building the right comparison. That shows why you are better than ResNet50, even though it’s not even for example, the state of the art anymore, but. Just to say that to improve something. And you need to have a very broad context in every, in each domain, in order to start before knowing how to implement these things in order to start differentiating between the things that are real progress and the things that more, look like progress than are real progress, or that the paper is not really proven that there is progress in, in that domain.
[00:18:25] And by the way, in the medical imaging domain, there are many papers that are written also by leading groups about deep learning and AI for medical imaging. And we find most of them not useful for our uses, because for example, they use very small test sets to measure their performance comprised of several dozens of scans when we are using at least thousands or tens of thousands of scans to measure our performance.
[00:18:53] And in dozens of scans, you can barely. Starts to, to understand the, the variability. And even if you improve the performance there, it really doesn’t reflect what you’re going to achieve in the real world.
[00:19:06] Gil Elbaz: Yeah, it’s definitely not obvious that given a very small subset of images, it will actually translate well to the population of the world.
[00:19:15] What has surprised you most in the last few years in the academia?
[00:19:19] Idan Bassuk: I don’t know if in the academia, but with relation to the academia, and it’s one of my main lessons about AI. And, I think that in retrospect, it’s not as much as a surprise. But I think that it’s an interesting notion.
[00:19:32] Today the narrative of AI and deep learning is controlled by the major academic organizations and they, by the major, uh, tech companies like Google and Facebook, Microsoft
[00:19:45] and these companies, most of the people also that, dominate the narrative are also academic in many aspects. And they work in research organizations inside these companies, which mainly publish papers, academic papers, and do an excellent job at that, by the way. And in parallel, the fact that the narrative is dominated by, uh, academic, experts also narrows it down a bit.
[00:20:10] I think to things that, uh, academic personnel more want to work on because academic personnel, they prefer working on things. I think that are less related to things that require operations, like building new data sets or monitoring your performance in production, et cetera, and the real world.
[00:20:30] And so this is the first thing. And the second thing is that having a really deep understanding of the product, for example, in many papers or other venues, you may see that they compare metrics such as the AUC, or the loss, while in practice, they don’t necessarily correlate to the value that your algorithm will eventually provide to the customer. So you might improve the customer value without improving these metrics, or you might improve these metrics without really improving the customer value.
[00:21:06] Gil Elbaz: There’s a big gap that’s created between what in theory is possible with these theoretical papers and what we’re actually able to bring to production in a straightforward, simple sense. And that gap is a very challenging engineering, expensive gap that needs to be bridged in order to bring this, value that is perceived in the academia to the real world.
[00:21:30] Idan Bassuk: Yeah, addition I think that the part of model development of improving the models based on the same data is highly important and it’s highly relevant in many cases to production and to the real world, but it’s only one approach. It’s only one direction to approach the improvement of your AI.
[00:21:48] For example, improving your data set, which is something that is almost not at all researched in the academia and building a larger data set and how you do it. And what are the methods to do it. Oftentimes it’s at least as impactful on model performance as, uh, improving the model based on the state data set.
[00:22:07] But most of the papers don’t even look into it. So there are many other aspects. For example, we have a data science team, which we call the AI Operations C enter inside our, inside my group and inside our company, a very strong data science team, which doesn’t develop AI models, but its role is to sit on the intersection where the AI meets the real world, where the AI meets the data in the wild.
[00:22:34] Once before we deploy the AI into each medical center, in the other respect doing the continuous monitoring phase, and you can see that data science in the real world on how the AI actually interacts and what data actually exists in the real world. There are many opportunities there that can sometimes improve our accuracy, accuracy.
[00:22:57] The doctor actually feels by dozens of percent and it’s something that it’s almost not at all explored in the academia.
[00:23:05] Gil Elbaz: So this is extremely interesting. I’d love to dive into more what this AI Center or AI Operations really means in your org, because it sounds incredibly interesting.
[00:23:16] I will note that I completely agree with you that the data aspect has been far overlooked by the academia. And I think that one of the challenges is, that there is kind of a bias where they don’t have access to this amazing data. Right. Right. You need real products in the real world to get real data.
[00:23:32] Right. And that’s something that the academia lacks. And so that’s really, I think one of the reasons that the academia is so stuck on improving these models, even though the incremental improvements are at least from what I’ve seen are substantially less effective than improving the data and adding more data and adding more variance.
[00:23:49] So it seemed to me a very interesting and unique the AI Ops team that you referenced before. Can you tell us a little bit more about what they do and what their goal is?
[00:23:59] Idan Bassuk: Yeah, sure. But maybe I’d love a bit before I’ll dive into the team, which is a fascinating thing. I’d love to maybe zoom out, to give the full context from which this team and several other teams emerge from.
[00:24:12] Gil Elbaz: Sounds perfect.
[00:24:13] Idan Bassuk: As I said earlier at Aidoc, I lead the AI group, the concept of the AI group. It’s not only a group of the algorithm, engineers or researchers, but it’s contains all the teams that are responsible in any way for developing the AI or bringing it to production to the real world.
[00:24:33] in parallel to this AI Operations Center, this group also contains several other teams each of them part of the concept of holistically attacking the AI challenge from many directions. So. We have a data engineering team, data engineering is a known concept in the industry.
[00:24:52] And even in companies that don’t do AI, but the data engineering team of the AI group is responsible for data engineering platforms. enabling data mining, enabling building the data sets and the platform which enable to develop the datasets across dozens of types of data and scans and reports and medical records, et cetera, et cetera.
[00:25:16] we already have petabytes of scans. You need platforms in order to find the needle in the haystack, So we have a data engineering team which builds platforms that enable us to utilize the data to the best extent possible, there is a notion, in AI, that more data is better, but more data is not necessarily better because most of the data in the real world, and I think it really relates to the things that you are doing as well. It’s not interesting. It doesn’t teach the algorithm anything new. Most of the patients are healthy. our goal, we all, we train our law on large data sets. But our goal is to train on the smallest datasets possible, which contains the most interesting scan and not on the largest data set possible. We think that that’s not the correct KPI.
[00:26:01] So that’s one team. Another team our group actually is the data set engineering or dataset development group. It’s a generalization of the concept that in many companies is just called data annotation, but we think that building a data set on which you train and test your AI, especially in the challenges of the medical world is much beyond the notation, which by itself is not simple in the world of AI, since you need a trained radiologist in order to do it.
[00:26:29] Gil Elbaz: It’s extremely complex expensive. Even if you have real radiologists annotating and for each image, it could be incredibly challenging.
[00:26:39] Idan Bassuk: Yeah. Yeah, definitely. But dataset engineering is much beyond the annotation of the scans, even. It’s how you use the platforms that I spoke about a few minutes earlier to mine, the most interesting scans. You need people who understand not only the medical conditions, not only the hemorrhages and where to find them, but also the data variability, the physical properties of the scans ,our customers’ needs, in order to choose the most important and most interesting scans to the algorithm.
[00:27:08] Another challenge is how you build the test set on which you test your algorithm. And they said it, we have very significant data variability, right? So we don’t just want to measure our accuracy in general. We want to have an understanding internally of how our accuracy is affected by different cross-sections of the data, by different parameters and aspects of the data.
[00:27:34] If you don’t build the dataset in a way that enables you to separate correlation from causality, you will never be able to do it even with the best algorithm engineers.
[00:27:43] Gil Elbaz: So just to touch on that, cause it seems like this is one of the most important parts of the computer vision development life cycle, in my opinion. So How do you guys build your test sets is it one giant test said, are they divided? Do they each focus on something? the way that you guys mine it is super interesting and I think is something that we’d love to hear more about.
[00:28:04] Idan Bassuk: These experts that know how to mine the scan. So we have several roles which are in charge of it. We call them AI Dataset Engineers and AI Dataset Analysts. So we actually hire algorithm, engineers, deep learning engineers, which are experts in deep learning and specialize on the data-centric, AI aspects of the algorithm development. to guide the data set development and the dataset engineering efforts.
[00:28:29] So about the test set, it depends on the phase in the project, of course, but they are typically in the size of tens of thousands of scans. it’s very, very difficult to do it by the way, because the prevalence of the medical conditions, they are very rare now in order to reach the size in conditions in which only one out of 100 scans will be positive to that medical condition, you need to be able to really find the needle in the haystack. The basic aspect is first of all, find enough positive scans and we developed natural language processing capabilities in order to enable us to, to do it. Because each CT scans has over a billion pixel, and to find each and every finding in which scan is positive in which isn’t, you wouldn’t be able to do it at scale.
[00:29:19] So we developed natural language processing capabilities, and we developed a natural language processing framework actually, which enables us to quickly develop an algorithm for each new medical condition.
[00:29:33] Gil Elbaz: can you break that down for us? I want to understand kind of the connection now between how does the NLP actually accelerate your development of a new algorithm?
[00:29:40] Idan Bassuk: Yeah, so we have a huge database of many, many millions of, reports First of all, we have data from institutions all over the world, many institutions that we have partnerships on the data together with them and enable us to use their data for the training purposes for the algorithm development purposes.
[00:30:01] And we don’t only have their CT scans or their x-ray scans. We all also have the reports that the radiologist has written in the hospital when the scan was analyzed in the past, in real life and summarizes what the radiologist has seen on the scan. And by the way, this report is not a two liner. “I saw a brain hemorrhage”, End. It’s a 2 A4 pages a report often that explains each and every, a fissure of the brain and in each and every vertebra. What’s interesting there, even if it’s not life risking medical conditions, very complex reports, unstructured, in medical language.
[00:30:41] Gil Elbaz: Yeah. It seems extremely complicated to have such a long report with so many variables. So many things changing, describing so many things,
[00:30:49] Idan Bassuk: by the way, different languages we have, we need to be able to work.
[00:30:52] Gil Elbaz: Oh wow.
[00:30:53] Idan Bassuk: With reports in all European languages and in English. Yeah. So and in hebrew,
[00:30:59] Gil Elbaz: it would be extremely interesting to see if there are substantial differences in the reports or the length or the style between these different languages and cultures.
[00:31:09] Idan Bassuk: Yeah. Even different English speaking countries, even the word hemorrhage, they spell it differently by the way.
[00:31:15] Gil Elbaz: Oh my goodness.
[00:31:18] Using all of this data together, you’re able to then hone in on the images that you want to annotate so that you can annotate the most important ones first it’s similar to triaged, right? Like similar to the product itself. It’s like a, pre-pro like a product in a product to find what you need to annotate first in order to improve the algorithm in the best way possible.
[00:31:41] Idan Bassuk: Yeah, definitely. And we don’t only use the reports to find which scan has a finding or which doesn’t, but also to focus on the most interesting findings, because often, although not always, these reports mentioned which findings were more difficult or more subtle. And since we want our product to reach very high levels of accuracy that can really have the potential to improve the diagnostic accuracy of the radiologists, we need to be very focused, not only on the large findings, but, especially in the most subtle findings. And the reports really help us, do it.
[00:32:15] Gil Elbaz: Extremely interesting. And with these same reports in the future, or maybe you guys are already doing this, would it be possible to mine additional correlations between various, maybe unknown aspects or aspects that haven’t necessarily been researched deeply in the medical space to see kind of patterns or things that might be very unique.
[00:32:36] Idan Bassuk: Definitely. And by the way, not only in reports, And I’ll just give an example one of our products deals with, pulmonary nodules, the detection of actually, uh, tumors in the lungs in everyday language, or lung cancer. in most, or a large portion of the CT scans, the radiologist is able to identify the tumor, but is unable to say if it’s malignant, meaning dangerous, or benign from the CT scan itself.
[00:33:03] So if the tumor is identified, the patient is followed up with additional CT scans every several months or years. And the tumor is compared how it looked and how much did it grow and how did it change its shape. And eventually in many cases, more complex examinations are prescribed to that patient in order to classify for the radiologist, without AI, to whether it’s malignant or benign.
[00:33:32] So since we have access to all these types of data, we don’t have to train the AI based on the CT scan independently and what the radiologist was able to say when he interpreted the CT scan in real time. We can start from the end. We can start from the medical records of the patients in our databases, see which patients were eventually diagnosed with a malignant tumor and train, especially on these tumors to enrich our data set with the most interesting tumors but also.
[00:34:05] Seeing the patients as were followed up many times and eventually classified as benign are also very interesting because these are exactly the cases where the radiologist wasn’t able to say it by himself. And we can use this information when we train our algorithm.
[00:34:21] Gil Elbaz: So in addition to NLP, which is one way you guys use the future in order to look back into the future.
[00:34:30] Idan Bassuk: Right?
[00:34:30] Gil Elbaz: Amazing, amazing. So yeah, we have a time traveling VP AI with us today. I’d love to maybe dive in now deeper into how to get these, capabilities into production. So we have this amazing pipeline of data analysis, understanding, gathering the data, triaging what data is important. now you want to train these algorithms, how do you go about that?
[00:34:56] Idan Bassuk: About training the algorithms themselves?
[00:34:58] Gil Elbaz: What architectures, you know, you have an infinite amount of architectures that you could choose from. You have a ton of academic research in many different fields.
[00:35:08] How do you go about finding what you need to use to train?
[00:35:11] Idan Bassuk: Yeah, so our method is very data-driven. We start from our baseline architectures because eventually there are many similarities between different medical conditions. So we don’t need to reinvent the wheel and start from the paper that open source code was published last week every time we, we start with the new medical problem.
[00:35:33] There are definitely huge challenges that we encounter in each and every new medical problem that we need to adapt to. But the baseline is, is overall similar. So we have a baseline architecture, which is a multi-stage, you know, CNN detection, segmentation framework.
[00:35:50] 3d of course, and By the way, even before training on the data, choosing the annotation method is critical and it can have a huge impact it can cap your accuracy if you don’t choose the correct annotation method.
[00:36:05] Gil Elbaz: Let’s dive into that. does that mean?
[00:36:07] Idan Bassuk: Before we even start training the algorithm, one of the most important steps in the data annotation actually in the data annotation method, when you’re talking about medical conditions, there’s not necessarily a single definition of which pixels exactly contain this medical condition in the image. There can mean many types of appearances or many types of indicating factors, and some of them can be relevant in some cases and irrelevant in other cases. Another consideration for choosing the annotation method. As we said earlier, the operation for annotating scans is very, very expensive. since we annotate very large data sets, it pays off to build algorithms to expedite the annotation work, right? So certain annotation methods might be theoretically better for your algorithms, eventually accuracy, but harder to build the tools that expedite their annotation process. And you will end up with a smaller data set, which might harm your accuracy eventually. So there are really non-trivial considerations
[00:37:07] for example, a blood clot in an artery, which can cause a stroke or lack of blood flow to your lungs and really life risking conditions.
[00:37:16] And we have several products in these domains. What do you want to annotate? The boundary between the area where there is blood and the area where there is no blood flowing anymore? Or do you want to only annotate the filling defect, the area, where there is no blood anymore, or do you want to annotate the clot itself?
[00:37:36] And in certain scans, this will be clearer. In other scans that will be clearer in certain scans, the medium between these two areas will be a very clear boundary. In other scans, it will be vague and not necessarily, you will be able to determine it.
[00:37:52] When you build many algorithms that detect blood clots in different organs, in different arteries, you want to be able to do the transfer learning between them. And you want to choose a single annotation method that potentially will enable all of them to benefit from this holistic datasets of clots in the brain and in the lungs and in the neck, et cetera, et cetera, because there is a lot of neutral features between them. And that’s another consideration in choosing your annotation method, then.
[00:38:20] And you need a deep understanding of algorithms and how deep learning works in order to choose denotation method. Because for example, if you choose to annotate the phenomena in a way that one of the appearances in the phenomena that is not necessarily a local appearance, very localized appearance, then you might be relying on information that is outside of the receptive field of the algorithm for example.
[00:38:46] Gil Elbaz: And if it’s too local on the other side, then you might get a lot of false positives, a lot of things that look similar, but it’s just because the very local area is very similar, but in a global more holistic approach, or with a larger receptive field, it would be known that it is either false or positive more easily.
[00:39:07] So this is interesting. So there’s no playbook for this, right. you’re writing the playbook for this.
[00:39:11] Do you guys talk with radiologists on how they perceive these things?
[00:39:15] Idan Bassuk: Of course, yeah. Both in these stages and in the later stages of developing the models and of course testing them as well. We have in-house radiologists in our company that we are in constant communication with, the overarching rationale, I think, first of all, not only in our company, but I think the overarching rationale in developing AI algorithms and AI models is to take a complex problem, but to steer the network, or the algorithm to break it down to more simpler sub problems, even if you train the algorithm end to end eventually, but the architecture points out and breaks it down to different sub-problems for instance, Like CNN’s did for images, right?
[00:39:57] You break down the problem of analyzing an entire image to analyzing first, very smaller receptive field, and then gradually growing receptive fields. And I think that in our company, the overarching rationale inspired by that is to steer the algorithm, to learn, to look at the data in simplified sub-problems like a radiologist would look at this data, right? So we are constantly trying to see, for example, look at our algorithms, false negatives and false positives, and understand why the algorithm made this mistake and why a radiologist wouldn’t have made this mistake and what data or what method the radiologist is using that we haven’t yet incorporated into our algorithm.
[00:40:43] Gil Elbaz: We talked a lot. How the annotations actually affect the bottom line. okay, so we have data, we have annotated data and enormous amount, And the question is, does this flow within your organization? How do they actually take this and then work on the algorithm? And in the end we want to reach production. Is there kind of a research phase or an initial phase, how does this work by you guys?
[00:41:06] Idan Bassuk: Yeah. Great question. So as part of the organizational structure that I mentioned earlier, we have two groups inside the AI group. We have the data department, which contains many of the teams that I mentioned from the data set engineering group to the AI Operations center and the data engineering team. And we have the AI algorithm engineering group, which contains the algorithm engineers and AI software engineers, which are specialize in engineering AI systems.
[00:41:35] And by the way, not inside my group, but we also have a product department, right?
[00:41:40] So from the earliest stages of the inception of the idea of the project, there is a project team which consists of a person from each relevant team in this organization which is meant to discuss the problem we’re still vetting whether or not we want to develop an algorithm in this area. And how challenging would it be to acquire the data or developing the algorithm? How much value will the algorithm that we expected, we’ll be able to build. will provide to the radiologist and in which faith.
[00:42:12] So that’s from the earliest stages, even before the dataset is annotated throughout the dataset development and engineering phases, in which the algorithm engineers are deeply involved together with the data experts on defining the methodology and the amounts and the data sources and the data characteristics, throughout the development phase, in which once a data set is fully ready, it’s just already in our database and it’s fully ready for the algorithm engineer to use it. So in that stage, we’ll already be having for several months, weekly status meetings of the project team, across all the relevant teams on the progress around the dataset development, you need to make many decisions around things that took more time than you expect and things that took less time and surprises that you found while annotating the data.
[00:43:06] So you already have a very tight coordination and the algorithm engineer that is going to develop this algorithm is already very into and, inside the project. then the transition to the development phase is very, very smooth. The engineer already is very deep into the problem and has been guiding these efforts for several weeks or months.
[00:43:30] Gil Elbaz: That seems like a really great way for the algorithm engineer himself or herself to actually be very, very connected to the end product. And that’s one of the most important things from my eyes, from what I’ve seen in bringing these kind of AI capabilities to production is that the algorithm engineer itself needs to be very close with the product needs to really understand what’s important and what isn’t important because at the end of the day, the reason, yeah, at the end of the day, the reason that a lot of really amazing capabilities don’t reach production from what I’ve seen is a miscommunication or misunderstanding between the product or what’s actually needed in the end and the algorithm developer or the data or the data annotators or the whole pipeline pretty much needs to be very much aware.
[00:44:18] Idan Bassuk: Definitely. And we had a case in which one of our, uh, algorithm developers, thanks to her being involved from the earliest stages of this development of this product and algorithm in a later stage of the project, we saw that our accuracy is not as high as needed as the required by product, but then she understood that we’re measuring the metric that we use to measure in similar products.
[00:44:44] But in this case, something a bit different is important to the doctors. So maybe we don’t need to improve the algorithm. We don’t, need to improve the dataset. We need to change the metric in order to measure the value of the algorithm in the same way that the doctor that uses this product will measure the value of the algorithm.
[00:45:01] And then with 10 lines of code, we saw that our accuracy as the doctor will perceive it is actually much higher than we thought previously.
[00:45:09] That’s great.
[00:45:10] Gil Elbaz: I’d love to understand your views on clean code and clean, quality code architecture around these AI products.
[00:45:19] Idan Bassuk: Yeah. Sure. So about building AI in production grade, I think the AI software engineers have a large part in it and I’d love to talk about it, but I think it’s also has a lot to do with how the algorithm engineers work. Okay. So I’ll give a few examples when we develop an algorithm, data variability is one of our huge challenges.
[00:45:42] And I think that it’s something that many people won’t expect that around 30 to 50% of the effort that our algorithm engineers are investing in developing value with them is in the stage of measuring the algorithms performance.
[00:45:57] Gil Elbaz: Benchmarking the data, understanding the quality.
[00:46:01] Idan Bassuk: Differentiating between correlation and causality of different features of the data and how they impact the accuracy of the algorithm, et cetera, measuring how it the accuracy on different medical centers and how robust your accuracy is and how variability is between different medical centers, what impacts it.
[00:46:18] So a lot of the work is measuring and evaluating the performance of the algorithm, and it requires an engineering mindset. It’s not only wanting to develop the model. It requires an engineering mindset. It requires really a passion to really work methodically and thoroughly. It’s something that many people that want to develop AI or develop AI don’t necessarily imagine themselves doing.
[00:46:43] And it also relates to culture and it starts from the hiring process. And we look for people who are very passionate about these kinds of things, and by the. More than possible to find people who will be the best researchers and also will be very passionate about the engineering of their, of their algorithms to the smallest details.
[00:47:04] it starts from that. for example, I said it, we differentiate between correlation and causality. how do we do it? We perform certain analysis of the data of the performance on the test. and eventually you have a graph and you want to look at it and make a conclusion out of this graph. So when we complete the project, and we see which of these graphs, and we have dozens of graphs and different parameters, et cetera. We see which of these graphs are going to impact our decision-making in production, how we deploy our algorithm We code review the scripts, that generated these graphs, the algorithm engineers not only develop the algorithms, but when they complete development of the algorithm, they also code review not only the code of the algorithm, but also the ad hoc scripts, that generated, these graphs, because a tiny mistake in this, uh, you know, pandas script can lead you to the wrong conclusions of, uh, the impact of different parameters on the accuracy of your algorithm, which is not less important than the actual accuracy of your algorithm.
[00:48:12] Gil Elbaz: It’s extremely important. Yes. And it’s amazingly, you go to that depth to fully check. I mean, it’s, it’s very, very important in the medical industry and I’m personally very happy about the emphasis on the, and the importance of testing the algorithms in the best way possible. And then checking that even the small things like the graphs themselves that you draw your conclusions from are correct.
[00:48:34] Idan Bassuk: Yeah. So, and the other side of the, of the same coin of the engineering, the algorithms is the software engineering team. So actually, if I can tell a small story I started in, in our company is an algorithm engineer, but eventually I found out that a lot of my work is actually doing the software engineering of developing the infrastructure that runs down with them and in research and in production, But it’s at least as important as knowing how to develop the models
[00:49:05] Gil Elbaz: It’s very challenging in many cases, from what I’ve seen to get people that are in the mindset of algorithms, algorithms, algorithms to a more software engineering mindset
[00:49:16] Idan Bassuk: Engineering focused mindset, for example, reviewing the graphs
[00:49:20] when we developed one of our first algorithms, we understood that the same person can have many different findings. Okay. For example, it’s not hemorrhages, but just to make it clear, the same person can have even 20 different hemorrhages in that organ.
[00:49:35] Okay. But the doctor, the radiologist, if it’s more than six, usually it’s not really interesting if it’s six 20 Okay. So we built a mechanism it’s even, it’s not inside the neural network. It’s in the post-processing phase to filter the smallest, findings, for example. at first.
[00:49:52] Before we deploy this algorithm to production. We found out that the algorithm engineer that wrote this code, I don’t remember. It might even be in me, actually sorted from the smallest finding to the largest, instead of from the largest finding to the smallest. So you can end up showing the six least significant findings to the doctor, and he can see a huge finding in another place and say, okay, you show me the least significant findings, but that’s embarrassing.
[00:50:19] He didn’t see this huge, like football sized finding. And, uh, eventually you need to develop all of the parts of your software in a very well-engineered way. You need to write tests. You need to work with design patterns that don’t yet exist in the world of AI software systems. and we discovered that before we deployed this algorithm to production, because we did poorly on our code. at that stage, it was really in the first year of the company. And we didn’t yet write tests for most of the things. We did actually write some tests, but today we are obsessive about writing tests my conclusion from that, 70% of our work of the algorithm engineers work became software engineering. And then 90% of our work became software engineering,
[00:51:07] by the way. And I personally loved it the rest of our algorithm engineers at that time, don’t think that they loved it, but luckily they were a perfect example of the value that we call ownership that we really champion this value in our company. They were just really happy to do whatever is necessary to get this algorithm to production. These guys are bright. Investing 50% of your time in measuring performance and the accuracy of,
[00:51:32] Gil Elbaz: So tell me how, do you get them passionate about this? How do you actually get them to the, to the point where they understand the importance and they’re, fully onboard with this software engineering first approach?
[00:51:43] Idan Bassuk: Yeah, so we, look for it during the hiring process. One of the things I believe the most during the hiring process is something I call extreme transparency. We really believe that all the roles in our company are, are amazing and fascinating. And we have really superstars in all of these roles. And in parallel, each role has its nuances and aspects and something that can be a huge advantage for one person can be a huge turn off for another.
[00:52:09] So we really believe in telling our candidates and we invest hours in each process in talking to our candidates and explaining exactly. How the role really looks like and explain the different aspects of the role. And I think that once you are really upfront with these things, people that are really not motivated by these things, they will eventually filter themselves out.
[00:52:35] And I look at it as a win-win situation because bringing a person, no matter how smart she is, just to eventually understand it, it’s not the role that she was looking for. It’s a lose, lose situation for both sides. So when you are transparent about how your role really looks like, and we invest hours in giving this transparency to each candidates, because we really think the hiring process, you know, it’s a two-sided process and it’s the candidates, right and it’s our, it’s our right to give the candidate this transparency. Then you really see which people’s eyes light up when you speak about these things. And even if they didn’t necessarily work at places. Previously that we’re so passionate about these standards to this level, you can see that their eyes light up.
[00:53:24] The people that we will choose are people who are very strong on the algorithmic side. And most of the people that we hired in the previous year were like PhDs and 10 to 20 years of experience from excellent places ,and in parallel, their eyes light up. And they were able to give examples, even if not all of the things and exactly how we look at the engineering side, but how they approach it in the previous places that they worked in. And they were very, uh, excited about joining a place that works like that.
[00:53:53] Gil Elbaz: We also, from the get-go, we have our head of engineering. or our director of engineering who was very, very focused on clean code, really from the beginning stages. And, very experienced guy has all these Python theories of how to do design patterns and better ways, and how to actually work in larger teams and able to merge more seamlessly.
[00:54:15] And he has all of these very unique kind of insights. And I’ve personally learned a ton from working with him. Definitely. But yeah, at a certain scale, it becomes paramount because every bug, every challenge that’s unforeseen is a substantial impact to the actual progress of the business.
[00:54:34] Definitely. And they can remain hidden for years. I think that a lot of it is where the AI software engineers come in and maybe I’ll, explain it a little bit about what they are doing.
[00:54:44] Idan Bassuk: So our AI software engineering team. Each algorithm that we are developing, you know, it’s not a completely different world. there are a lot of commonalities between our different algorithms. We have different types and we have the classification algorithms and we have the x-ray algorithms and we have this detection, but there are several families of algorithms. since we have many commonalities, we are able to focus our software engineers on problems that are broad rather than ad hoc.
[00:55:16] The role of our AI software engineers is not to okay. Build me a Docker and a REST API for this, uh, algorithm. And in two days I’ll have a new model and build me another REST API for that algorithm. But it’s how do you take this family of 10 algorithms at scale of 500 hospitals? And you build the infrastructure to do the same thing in 2000 hospitals or to 10 X, what you’re doing in 2000 hospitals.
[00:55:41] The first area of responsibility is building the infrastructure for research and development of the algorithms. This is the internal infrastructure that the algorithm engineers use when they run the experiments.
[00:55:53] But it’s not only the cloud instances, but for example, as I said, we have different types of algorithms from x-ray MRI, CT detection classification, NLP.
[00:56:06] Do you want this completely different code base for each of these types of algorithms? And when you are doubling your algorithms team every year, you want them to each new algorithm engineer to be able to learn a completely different code base for each new type of algorithm, you will want him to develop.
[00:56:21] It’s less scalable and it’s less flexible. So you want to develop the algorithmic infrastructure in a way that is as close as possible to a single mental model for all of these types of algorithms. That from the one hand streamline the work of the more common use cases.
[00:56:39] But on the other hand provides the flexibility the researcher needs when she, uh, improves the model. So in order to do that, the first thing that they do in the company is do the famous, uh, Stanford CS231n course, the deep learning computer vision course, that’s the first thing they do in their training the AI software engineers, because they really need to have a deep understanding of the algorithms in order to find the right abstractions and the right design for these systems.
[00:57:07] That’s great.
[00:57:08] One of the challenges that I see many teams having is reproducing their models and you develop a model, it works well and you take it to production and it works.
[00:57:17] And then you want to improve it after six months, for example, you want improve it, but you’re not able to reproduce previous model because the data change because your code changed, because your code breaks, there are 10 scripts and no one is sure which of these scripts were then Yeah. And for example, another responsibility of these teams is to build continuous integration and continuous delivery, CI and CD for AI. Okay. So for example, every algorithm that we take into production and prerequisite that we have before we take it into production is that we have a CI process that runs on a weekly basis, that we, the most updated version of our database and, of our code, et cetera, trains this algorithm from scratch and proves that it’s able to reproduce the same accuracy that we have demonstrated when we deployed it to production.
[00:58:08] Gil Elbaz: Or better.
[00:58:09] Idan Bassuk: Or better. Yeah. And once we do that, we are sure that when we get back to developing this algorithm, since this test is running every week and it works, the only gap for continuing to develop this algorithm is taking the script and using it as the basis for your development. And you have everything there, the data and the code and the hyper parameters etc.
[00:58:31] Gil Elbaz: regarding the data, I think that there is a challenge there, right? Because if you version the data and you have an enormous data set, then it it becomes a challenge. For instance, if you upgrade the annotations, do you keep every version of the data or do you have some kind of system for, for managing the deltas there?
[00:58:48] Idan Bassuk: Yeah. Yeah. It’s a great question. And I think it really depends on, on the use case. And in some cases we have a system that we use for data versioning and in other cases, the impact of version, the data exists, but it’s not large enough. So. Yes, dedicate the efforts to build the infrastructure around it.
[00:59:06] And overall, I really think that this example of being able to reproduce your algorithm, the infrastructure around it, the way I see it is a prerequisite for working agile in AI, because what is agile eventually agile or lean, lean startup for those? No, it, what is agile? It’s working iteratively its First version of your algorithm,
[00:59:31] You don’t want to develop the algorithm that you will have five years from now and only then ship it to the customer. You want to get feedback from your customer on premature version to see what your customers think should be improved the most. But if you ship it to the customer and get feedback, and then you need to invest several weeks or months, just to be able to continue where you left off, it’s an impediment to being agile in AI, because it will encourage you to make longer iterations because resuming is more painful.
[01:00:04] Gil Elbaz: Definitely. Resuming can be extremely painful if you don’t have the infrastructure in place, the mindset in place when you’re developing the original code. Yeah. It’s definitely one of the challenges that we see, especially with smaller companies.
[01:00:15] Now, the ability to work agile in AI is very non-trivial today.
[01:00:20] Idan Bassuk: this specific aspect of agile of working iteratively, I think that this is a highly important pattern not having impediments that encourage you to make the iterations longer than what is make sense from the product perspective.
[01:00:35] Right. But I think that, uh, from other aspects of, of agile or frameworks, like scrum, like, uh, sprains and what is the length of them and et cetera, I think that generally it’s a good idea to take the framework of, of agile and scrum or something similar to that, but not stick to it too tightly, by the way, not necessarily only in AI, but in general it will be adapted to the needs of your organization.
[01:01:04] Yeah. And to the integration lengths that make sense and to your team size and the types of people in your team and what makes sense. And, uh, in order to do it, I think that you need to be very experienced in managing these kinds of teams and you need. Really deeply understand the principles behind these frameworks in order to make, to really adapt it well.
[01:01:26] And not only thing that you’re making a right is a good decision, even though you’re doing it, just because you maybe didn’t understand what the poet meant when they talked about doing it in the set in a certain way.
[01:01:37] Gil Elbaz: Of course. it’s always something that should be fine tuned, right? Aidoc today and Aidoc in three years might need different methodologies, different, types of teams interacting in different ways in different cadences.
[01:01:49] Idan Bassuk: Completely and Aidoc a year ago, worked completely differently than Aidoc today. And even today, different teams in Aidoc work completely differently than, than each other.
[01:01:58] For example, a maybe a year and a half ago, we would work like many other startups where, the algorithm engineering team and the software engineering team, which I managed at that time, we weekly meetings with the heads of the product department. Where we would prioritize the tasks. Right. And we have top-notch product managers and they are a fountain of good ideas. every week. And I think that it’s, a phenomenon that many people experience, they would hear what we’re working on and just ask for something small from the side and just another task to do in parallel.
[01:02:29] And then we would finish the quarter and we would ask ourselves, wait, but why didn’t we finish developing half of the number of algorithms that we plan to develop in this quarter?
[01:02:38] And one of the main answers for that was that we didn’t have enough control of understanding what goes in which additional tasks go in from the side and how they would impact the end result.
[01:02:50] So when we had this notion, we decided to switch to do quarterly plan. We plan each quarter in advance and which algorithms we’re going to work on, how much time generally, roughly, it’s going to take and who is going to work, and when. And it’s not, instead of being agile, it’s not in, instead of being a startup, you know, and working like a startup, you can still have the meeting with product and they can still want to get in tasks from the side.
[01:03:20] But now you have a baseline plan. And when they say, okay, I want to do this and this task that wasn’t planned in, tell them, okay, let’s choose who does it. And I can show you on the expense of what it will come.
[01:03:33] Gil Elbaz: Exactly. What does this impact at the end of the day, because there’s no free lunch people.
[01:03:38] Idan Bassuk: Right.
[01:03:38] And you can show them how much delay we already had in this project and a project that we plan to complete.
[01:03:44] And they wanted us to complete at certain months, we were already in a month delay and this will cause us another one to three week with. And they can decide for themselves because eventually product are the ones that need, in my opinion, or at least in our company are the ones that make the prioritization decisions and they are the ones capable to do it.
[01:04:03] And they can decide what is better if they prefer to postpone the project that was planned and do the new task or forget about the new task. And we still have unplanned tasks.
[01:04:15] Gil Elbaz: Do you guys set time aside for the uncertain
[01:04:17] Idan Bassuk: in certain teams certain, but not always.
[01:04:20] Yeah, not always depends on the team and how much we expect. we are still working like a startup and we are still. Performing unplanned tasks, but the amount of tasks that eventually are prioritized instead of the planned tasks is at least an order of magnitude smaller than what we had previously. So you asked about buffer, right? So do we leave buffer for unplanned tasks? So from my experience, by the way, buffer is it’s sometimes it’s good, but sometimes can also be dangerous. Why? Because a buffer could be a backdoor for tasks that wouldn’t really stand the test of prioritization against the other important tasks.
[01:05:00] Right? If you have a backdoor to insert these unplanned tasks just because there is time dedicated to them. It doesn’t mean they are more important than, than other tasks by giving them a buffer. You made the decision to slow down your other efforts. So why not just prioritize them head to head with your other efforts?
[01:05:19] Gil Elbaz: To wrap this up, I have one last question and that’s really for new people coming into the field, people that have just finished their second degree, let’s say in computer vision or people that have just finished their first degree and want to be integrated into either the software engineering or the algorithms side, what is your number one recommendation?
[01:05:39] Idan Bassuk: I personally think is that the most important thing for a junior engineer, someone relatively new to the industry is not what project you are working on or what company you are working on, or what domain you’re working on is the quality of the people that you will be working side by side with.
[01:05:59] And by the way, even if you will will not necessarily do the most sexy tasks in this team, but you will have the opportunity to watch experienced engineers and how they tackle the complex problems and discuss it with them.
[01:06:13] I think you will learn much more from it in many cases. As talented as you are, I think it’s very valuable for most people to have someone to learn from.
[01:06:23] Gil Elbaz: And you’re talking both about mentors and about the people on the team.
[01:06:27] Idan Bassuk: Yeah. Not necessarily your formal mentor. I think that just the people on the team having people like this can really change your pace of growth
[01:06:36] Gil Elbaz: Definitely. I agree a lot. Yeah. That’s great advice. And thank you very much, Idan for joining us today. It was truly a pleasure. Thank you.