Cloud Native Computing Foundation Kubernetes AI Day EU 2022, 19 May 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Kubernetes + AI Joining Forces in the Battle Against Cancer - Wojciech Małota-Wójcik, Ridge

Description

Kubernetes + AI Joining Forces in the Battle Against Cancer - Wojciech Małota-Wójcik, Ridge

A doctor is able to treat one patient at a time. On the other hand, engineers may create software analyzing thousands of cases every day! This presentation will focus on how geographically distributed Kubernetes clusters and AI are indispensable tools for IT professionals and computational biologists as they join forces to battle cancer. Computational biology — combining AI, medicine, mathematics, statistics, IoT and cloud computing — is increasing the precision and capacity of diagnostic processes. This evolution requires petabytes of storage, low-latency networking, efficient GPUs and ease of deployment on a massive scale. This leads us directly to the growing need for highly geographically distributed Kubernetes clusters, running as close to the hospital as possible. The presentation will review the basics of processing cancer images and then show actual examples of how AI algorithms are developed and deployed on K8s clusters and used by doctors to perform life-saving treatments.

A

Hola hola everyone nice to see here so early in the morning, so let let's start the party welcome people in the room, the ones who watch us on the internet. I hope next time I will see your smiling faces. It would be nice as I've, as the guy already said, I'm voice ma otavoichik. I work in rich company, where I'm a proud member of great development team working on distributed cloud services after hours.

A

I work on my phd project, analyzing cancer images and trying to find some patterns in them, which would lead us to inventing better treatment methods.

A

Today, during my presentation, I will try to connect, though, to those two disciplines and will take you on a fascinating journey inside the human body to meet the greatest enemy of our times. Cancer.

A

Do if you are old enough, oh it doesn't work.

A

If you are old enough, you may remember once upon once upon a time animated movie when us when I was a child, I watched all the episodes all the time and you see what I do today so be careful. What do you what your, what your children watch in tv- and this is the exact place where we are going to today uh to meet the world of ourselves?

A

So please open your minds, fasten your seat belts and we are going to be 800 times smaller now to visit the fascinating world of our body.

A

So this is the slide prepared by taking taking a slice of resected prostate, and you might see all those darker violet regions there, which are the fragments attacked by cancer. This is the very advanced stage. This patient, probably don't live anymore, but if you are scared that you are sick, don't worry. This feeling will go away tomorrow and- and this is what we are going to analyze with ai models.

A

So let's go further. Now we are 400 times smaller and here on the left, you see a fragment of cancer which is developing it started by it started from a single cell, reproducing like crazy and on the right. You'll see much more developed fragment of the image where everything went wild. You see that cancer is everywhere there before going further with analyzing those images.

A

Let's talk a bit about our goals, usually when we work with cancer images trying to find some treatment methods, we have two groups of patients being treated some way and they present different outcomes. Some of them are healthy after the therapy. Some are not, and we would like to predict this therapy outcome. This treatment outcome by looking at those slides taken in advance before before treatment is applied by the histopathologists.

A

So this is our goal. This is very difficult, so yeah, the histopathologist is responsible for choosing the right therapy for the patient. The patient takes time is harmful to our bodies and costs money.

A

So it is very essential decision to apply the right treatment as soon as possible.

A

If it fails, then python probably will die, because there will be no more time or even there will be no other therapy to apply at that stage.

A

So we research, so histopathologists are like neo trying to pick the right pill for the patient, and we researchers are like morpheus trying to get some insights of better decision may be taken which pill. Do you choose red or blue? If blue you may go home if red will continue so anyone, nobody, okay, so red.

A

So let's follow the white rabbit. Then now we find we landed and we are 800 times smaller. Now this is what histopathology sees in the microscope.

A

Basically, this. This is not how our tissue look like. In reality, this image is a result of applying two dyes, hematoxylin analyzing to our tissue hematoxylin, which is visible as darker regions. Here here right here here here.

A

Attaches to phosphate groups, building our dna inside cell nuclei and the algin visible here as pinky regions between those nuclear nuclei attached to other objects in in the cytoplasm of the cell. So by applying the staining protocol, we are able to visualize the tissue.

A

Let's start applying some mathematical magic. Don't worry, no calculus involved. Only linear algebra here in the center is the original image and on the left you see on the left and on the right. You see the transformation of this image, because each pixel represented by rgb values in the center may be transformed by solving some equations to other two images, one representing the amount of using attached to the image at particular point and second, one on the right representing the hematoxylin.

A

So, as I said, hematoxylin attaches to uh to the dna inside our nuclei, so you vis, you see clearly that nucleus, the the nuclei are visible here, but not here here. These are just dark points, because there is no hematoxylin in the project I want to present. Today we will just focus on nuclei, so basically we may ignore the aerosol image and we will focus on only on the hematoxylin one.

A

So another piece of information we need before we may process those images using ai, and actually this is ai.

A

The thing we need is to find the exact places where the nuclei exist in the image. This is done by using an open source neural network called stardust, which was trained to find new clay in the tissue images. So you see that it is very accurate, nice neural network, open source. Anyone may use it.

A

Having those two information, we may transform hematoxylin image and just ignore the pixels outside of the no clay, because, basically, we will analyze the the texture inside the nuclei, so everything else may be ignored.

A

You may see that this image still must be pre-processed before we may apply some some ai model. On top of it, you see that some nuclei are brighter. Some are darker.

A

If we process with the further with this image, our neural model may be trained to recognize if the nucleus, darker or brighter, it may focus on this feature, and this is not what we want to so before, feeding it into neural network. We want to equalize the contrast of each nuclei of each new close individually, so you see that they are much more similar now to each other.

A

Okay. So, let's start building our ai model using this image. The approach I'm proposing here is to classify each image inside the nuclei by taking 11 by 11 neighborhood around this pixel on the right. You see some examples.

A

You see that we we selected, for example, I don't know this- this pixel we have. uh We have this filled around of it and based on it. We will classify this pixel in the center of the image and we will do it pixel by pixel everywhere we will do it by using auto encoder.

A

This is very simple, but still very powerful, neural network architecture, which is an unsupervised method, because this is what we want to do here. We want to find some hidden knowledge not available to the histopathologists at the moment. So we look for unknown patterns. We will train autoencoder. This is very simple structure.

A

You just feed image to its input and you train it to get the exactly same. Outputs.

A

The magic of this model is made by introducing this bottleneck here. You see that the number of neurons here is much smaller than the number of inputs and outputs.

A

By doing this, we force this neural network to generalize generalize the information and find some patterns which might be useful in our case. As I said, we fit here 11 by 11 images, so it gives 121 assigned 16 bit integers on the input, and here we encode it to 10 60 64 bit float numbers, let's focus, let's, let's focus a little more on auto on the ah one more thing. Sorry you'll see that this neural network may be divided into two parts. One is encoder.

A

Second, one is decoder, let's start by by discussing encoder in more details.

A

So, as I said, we, inter we feed 121 pixel image on the input and we get 10 64 float numbers which gives us in total, 242 bytes on the input and 80 bytes on the output.

A

So, oh sorry, so you immediately see that this image, the information provided by the image, must be compressed in some way, because we neural network must return. 767 percent, less information by counting the bytes, so this gives out this is these: are the advantages of the of the auto encoder neural network. It compresses the data, it generalizes the knowledge and it reduces the noise.

A

Of course, it also introduces error in our case, as I measured it after training, the error is 4.7 per pixel in average, which might be understood as 12 grey levels in 8-bit space.

A

Let's go next here are some our examples of the images on the in each pair on the left. You see the original image and on the right, there is an image processed by both encoder and auto encoder.

A

So you see that some details are lost.

A

Also, images are smoother, so we will see how it how it behaves during the classification of the images and now, let's focus a bit on the decoder decoder is not the very useful part of the auto encoder architecture.

A

It is used only during the training and basically later we just remove it, but it provides one particular interesting feature.

A

Normally, the coder receives its input from directly from the auto encoder and returns the original image, but we may provide some random numbers to its inputs and then in response we will get a synthetic image taken out of thin air from time to time. In some projects it might be useful here, not really, but I prepared uh two videos for you presenting the presenting the updates returned from the decoder when I applied, applied, constant values of on the inputs and was progressively changing a single of them from zero to one.

A

So this is how the image changes. It may be a sign that neural network found some patterns, but you also see that the images we get here are very different from from the ones I presented earlier.

A

So who knows this is neural network right. We don't know how it thinks. Really so after taking all the pixels from the alden clay in the image and passing them through the encoder, I get points represented by 10 features right, 10, 10, float numbers. What you see here is the result of the umap algorithm.

A

Its goal is to transform high dimensional data to a smaller space at 2d in this case, but the important feature of this transformation is that it preserves clusters, so similar points are close to each other. So you see that we have some group here here here here right, so eventually it means that some groupings were found by the neural network, which is quite interesting.

A

Now, let's classify it, I took the simplest approach. I just applied a simple k-means algorithm to divide those points into eight classes, and this is the result, so you basically see that k-means eventually found found clusters presented by the umab. Of course, there are some differences, but this is the real world.

A

This is how the world works, and now, let's apply this classification to the original image.

A

You see that each pixel of the no clay is now marked with corresponding color. You immediately may see that software developers life is a pain, especially if you are ai developer because, as you might see, all those classes, blue, orange, dark red light, blue violet are, and of course, yellow are the same everywhere right.

A

So basically, our auto encoder learned how to recognize borders of the south, which needs more work to be fixed in this procedure, but we are not completely lost because we have two really interesting colors here, pinky and green, which looks like they classified the internals of the cell. Somehow, eventually we could take those two groups of the patients with different outcome of from the therapy and compare the histograms of the pinky and green regions run some statistical t tests, which is boring.

A

So I I'm not going to present it to you here and eventually we may find that if you, if we are very lucky that those categories may present some differences between two outcomes- and maybe maybe we have saved some lives here. Who knows- I haven't tried yet, but I will- and you see that that more work has to be done here to eliminate those border problems here.

A

So I presented some basic techniques and ideas behind presenting the behind recognizing the cancer. This is really very simple approach. There are much more complex things things going on, but of course we have no time to describe them here.

A

I hope you I got your interest, we need such people who, who love to do love to do such things, uh and let's now connect uh connect uh this approach somehow to kubernetes right, because this is why we are here today.

A

So this is how the archives of the hospital usually look like a lot of drawers dark room in the basement. Nobody wants to go there if you want to find something there, it's better to ask prosecutor, because basically, he is the most powerful man or woman to get data out of there if something goes wrong in hospital.

A

But this is what we researchers researchers must deal with. If we need data, sorry, the cancer, the cancer samples are stored there in analog way. These are just fragments of tissue on a piece of glass.

A

So if we need those data for training our models, we must go to the histopathologist asking for those samples. This histopathologist must go to archivist, find those samples scan them, and only then we may use those data for training our models so to make progress.

A

Situation must change, because histopathologists are extremely busy people it takes months until we are able to get our data for research, and it is never ending story to make her yeah and, and also it is hospit. It is important that in hospital r d department is not the first player, because hospitals are focused on treating on treating patients, so we just we just manage the situation. Somehow data must be accessible at any time by anyone, without bothering other people. It means that we must digitalize them.

A

Some medical images, like x-ray resonance tomography, are already stored in digital form, because this is how the modern hardware deliver them. Nowadays, as I said, the histopathological slides are still stored in analog form, because our organs still exist in real life and not in a metaverse.

A

So before we may analyze them using computers, they must be scanned and digitalized sooner so some hospitals already started investing in storing those slides and scanning them on a massive scale. Sooner than later, all the hospitals will follow this path.

A

So the current situation- sorry most of most of those hospitals- will migrate to the cloud, and this is the big opportunity for cloud providers for it. Specialists devops, whoever.

A

So both doctors and researchers will be able to access them at any time, but it will require petabytes of storage and low latency fast networking before between hospitals and data centers, because.

A

Because we need something better than providers offering data centers scattered across the country being far away from the hospital world is evolving and providers offering geographically distributed cloud services entered the market.

A

It means that data centers are now being located much closer to the hospitals offering shorter access to the end user.

A

Health related data is the most restricted type of information by gdpr. So all also legal stuff must be taken into account.

A

Cooperating with a cloud service provider living in the same country makes it easier to comply with applicable law.

A

Some hospitals taking r d seriously will go even further and build their own private in-house data, centers, so r d, department, storage and computing units are uh are stored, close to each other, offering offering faster access times faster model buildings and faster development.

A

This is how the hospital of the future, together with its r d department, look like there is a dedicated infrastructure for storage for storage, so both histopathologists and software developers may access the data at any time.

A

This is a perfect use case for kubernetes, because this is how we- because this is how clusters may be created dynamically, adding all the required resources on demand.

A

If you need reliable partner delivering geographically distributed cloud services with a list of proven successful deployments, please visit ridge.com contact us and ask and ask we will help.

A

So here are some reference links to the open source, libraries, open source models, even cancer images uh shared by the government on of usa. You may experiment with it. You may play with it whatever you like, and thank you very much for enjoying my presentation. If you have any questions, please ask anyone.

A

Okay, if there are no questions or are there any, do you have any questions?

A

Okay, anything in the slack we had somebody asking they could ask complex, but not actually asking questions. Okay. Thank you very much for attending my presentation. I hope it was interesting. This is the future actually right right, because you know there will be no more histopathologies training. A histopathologist takes 20 years, something like that right and we need more and more smart people and they won't be delivered.

A

So actually, the best investment we can do at the moment is to use very knowledge and experience to train computers to do their job because it will be cheaper and more effective and available for more people. So thank you very much. Enjoy your enjoy the cubecon today. Thank you.