Be ready to pick your jaw back up as we welcome 17yearold Brittany to the TEDxAtlanta stage. Applause Allright. So. Ok. That’s loud. I want you all to picture a little kid in the why phase. You know that kid that’s always asking questions. Why is the sky blue Why is the glass green And about million other questions per day. Now imagine, hypothetically speaking of course, what it would be like 15 years down the line if that kid was still in a why phase. As you may have imagined this isn’t quite a hypothetical situation.
I am a 17yearold who loves to ask questions. But, it wasn’t until I found science that I found my answers and more questions because with science the more you know the more you wonder. It’s an infinite process. One which I am completely enamored by. So, I love science because it gives us the opportunity to revolutionize the world around us. I think I first started seeing how many implications science really has where I was little. My brother was born premature and he spent over 60 days in a hospital his first year and many times after that.
We would go back and visit. I really grew to idolize men and women in scrubs because I could see that they were saving lives and using science to have a positive impact on the world. Then as a 7thgrader my passion for science grew even further. I was taking an elective on futuristic thinking. Quite by accident I stumbled across artificial intelligence. I was enthralled for computers can do things that humans can’t. That just is mindboggling. I went home, I bought a coding book, and I decided that that’s what I was going to focus on.
Global neural network cloud service for breast cancer detection Brittany Wenger at TEDxAtlanta
Now, this was pretty naive. I never actually coded anything. So I started with all the stereotypical beginner programs. Helloworld, card games, et cetera. And, eventually I did program artificial intelligence and it played soccer because at that time and still today I was a very avid soccer player. Now my passions have surpassed my grade school aspirations. I’m combining medical research and computer science to improve breast cancer diagnostics. And I did this so that my computer program could answer one simple question. Is a breast mass malignant or benign.
And this is really important because fine needle aspirates, the least invasive form of biopsy, are actually the least conclusive. This is a big problem because a lot of doctors can’t use them. If they can be used it would lead to earlier detection, less invasion and less cost. So, I tried to create a tool for doctors to use so that they could diagnose these fine needle aspirates. And I did this by creating an artificial neural network which is a type of program that can actually detect patterns that humans can’t detect.
I, then, put my artificial neural network that was diagnosing fine needle aspirates in a cloud because the cloud is this amazing elastic entity. They can scale to support usage by every hospital in the world. The current network is working really well. It is 99.1 sensitive to malignancy. And this is huge because this is the number that could save a lot of lives and can make the program hospital ready. In addition, I’ve run 7.6 million tests and I’ve proven that as I get more data the success rate should increase.
While the inconclusivity rate should decrease. So, why is this important One in eight women is impacted with breast cancer. And, these statistics are just startling. And, unfortunately, they are on the rise. However, when you have a personal encounter with a breast cancer that’s when the statistics become more of a reality. For me, that happened in my sophomore year of high school. My cousin was diagnosed with breast cancer. And, I saw the pain that it inflicted on her and the rest of the family. I was inspired to make a difference.
And my way of making a difference was to try to improve breast cancer diagnostics. So, I talked to you a little bit about fine needle aspirates. You know that they are the least invasive procedure. You know they are also the least conclusive. So, specifically in the United States, they are very rarely used. Most doctors won’t use them. What they actually do is they cause the patients about the level of discomfort of a blood test. A doctor sticks a fine needle into a palpable breast mass, and a few cells are extracted and then stained,.
And then, doctors look at them under a microscope. And then, traditionally doctors would decide whether the breast mass is malignant or benign. Now these are great tests because they are the most accessible to the general public, so they can lead to earlier detection. They are also about 4 times cheaper than the core biopsy, which is the current means of diagnosing aspirates. In addition, they cause women less physical and emotional scars, which is huge when you have to go through something as big as breast cancer. So that was the biology side of the project.
But what’s great about science right now is a lot of the breakthroughs are coming through interdisciplinary research. So here is a little bit on the computer side of the project. Artificial neural networks are actually programs that can model the brain’s neurons and innerconnectance to detect patterns that we can’t. And, this give them infinite potential because they are not limited on what we know and for something like cancer they are especially applicable because cancer is constantly mutating and constantly transforming. Neural networks are constantly learning so they can pick up on these changes learn how to handle them,.
And still diagnose the masses correctly which gives them a lot of potential. And not only are neural networks being used in medical realm, but they’ve got a lot of exciting applications. Currently they are being used at CERN where the Higgs Boson discoveries continue. They are also being used in more common place things such as your iTunes and Netflix all the suggestions are based on your prior experiences and they have neural networks figuring out what they think you’ll like. Future implications could lead to smarter rovers and a tutorial game.
That just got smarter as a player got smarter. So it lasted forever. Or even advances in earthquake detection. The other part of the computer science behind this project is the cloud. And a lot of us have probably heard about the cloud. It’s this huge technology buzz term right now. And what cloud allows for is that allows for servers to host my project. So what I mean by that is right now I am the only one and a few other hospitals are accessing my program. So we don’t use that much server space.
But tomorrow if a million hospitals decide they are interested in using it at the same exact time, my program will expand to all these different servers and I will be able to accommodate that. So I built it as a cloud service. And what that means is essentially my program exists out in cyberspace. And it’s just waiting for somebody to use it. So it’s looking for these messages and right now you can call my program via a web application. So you can go online to Cloud4Cancer.appspot. Doctors can use it. It’s working great.
However in order for a tool to be reliable and the whole purpose of this is to provide a tool for doctors it needs to be accessible. So some doctors are still using old PC systems. Others have moved towards mobile tablets. And there are new technologies that will emerge in the future. I can code platforms for these specific technologies. And they will be able to call my app and it will run. And one of the working examples of this is I am actually working with an institute in Italy.
They were able to create a program to reclassify the samples they already have based on my program’s inputs, and then call my service and get a response. You might be wondering at this point how the neural network actually works and how it applies to breast cancer. Well, the way my neural network works is that it takes in different inputs and decide on logical inputs from these fine needle aspirates. So doctors rate the inputs on a scale of 1 to 10. 1 being this input looks like a very benign attribute..
10 being this attribute would lead us to think this is cancer. So one of the examples of the inputs is Clump Thickness. Is a mass mono or multilayered Because monolayered masses is more indicative of mass being benign. Whereas multilayered masses are a lot more indicative of cancer. Another example is Marginal Adhesion. And this is going to quantify how closely the outside cells on the epithelial tend to stick together. If they stick together really tightly, then that’s a lot more indicative of a benign mass. Then another example is Bland Chromatin.
Which quantifies the texture of Chromatin in a nucleus. And if it’s fine, that’s a lot more like a benign mass. Whereas coarse is a lot more like a malignant tumor. So you might be wondering if doctors can look at attributes and they can decide which attributes would be indicative of a benign versus a malignant mass why did they need a tool to diagnose this And the answer is these fine needle aspirates aren’t cutanddry. This picture shows a benign mass. And you might have guessed that because the cells are about the same size and about the same shape.
However if you look at the cells, they have little dots in them. Those are nucleoli. Nucleoli should be barely visible and they should only be one per cell. These nucleoli are very prominent and there are multiple ones per cell. That would be an indicator of cancer. This next slide is also benign. Marginal Adhesion for this slide is great. Which will make you think maybe you don’t need to worry about cancer. However, you can see parts of the mass are actually multilayered, which is more of a malignant attribute.
And this final slide is cancer. And the cells are pretty spread out. They’re also different shapes, different sizes. So that would all lead you to think that it was malignant. However, some of the cells are actually devoid of their cytoplasm. If you look closer you can tell that. And that would be more of a benign attribute. So as you can see, fine needle aspirates are not easy to diagnose. There’s a lot of room for gray error. You might think that I just picked out three samples that were the least conclusive.
However, these are normal FNA’s. They are not the inconclusive ones and they are pretty indicative of the entire data set. So the way the neural network actually works is that it feeds in those nine inputs that are quantified on a scale of 1 to 10 and then it converts them into the artificial input layer. And this is something that’s really novel to my program. What it does is it transfers the number into the binary representation of that number. So now it has 4 input nodes. And what’s really cool is since neurons in the brain are either firing or not,.
And binary numbers are either ones or zeros, the digital spike and the neural spike are a lot alike. And bear in mind the whole purpose of this program is to replicate the brain. So this helps the network get success. The artificial input layer and the hidden layer then make 216 connections, or synapses. What’s really cool is the way we model the way brain communicates is actually via math. So what happens is that artificial input layer is then multiplied by a corresponding weight matrix. And then, the weighted inputs go into each of hidden nodes.
Are summed up via summation function. And then finally they’re sent to the sigmoid activation function. The sigmoid function is a logistic Sshape curve. And essentially it’s going to convert the number on a scale of 0 to 1. One being This node is definitely on. Zero being This node is not on at all. A similar process then occurs between the hidden layer and the output layer. And the whole purpose of this slide is to show you that through math we are able to replicate the way the brain thinks.
This is a backpropagation neural network meaning it’s going to learn based on its experiences and mistakes. So these connections are constantly updated once it learns what’s beneficial. So this is the sigmoid curve that I was talking about. Something that’s really unique to my program and really important and instrumental on its success is the fact that I weigh malignancy heavily. It is really important to diagnose cancer patients correctly. So on a scale of 0 to 1, Ix27m only going to call mass benign if the network thinks the value is under.2.
Programming an artificial neural network is not an easy task. And, it’s one. Laughter So I’m not afraid to admit I actually failed twice before the successful network. The first time there were more errors than code. So I ended up scrapping the entire program. And the second time I compiled I was really excited. I started running my tests. It’s actually worse than flipping coin in diagnosing breast cancer. But the important thing about science is you learn just as much from your floped experiments as you do from your successful ones.
So I was able to take what I learned from my experiments, and improve upon that in this third implementation. That’s why the neural network gets more success than previous university trails, and more success than its commercial counterparts. I am working with raw data. So the University of Wisconsin published this data, public domain on UCI Machine Learning Repository. I’m using all 681 samples. Ix27m not stripping any outliers. And I’m not having doctors go through and weigh input importance. I’m letting the neural network use its own brain to figure that out.
I also have this artificial input layer which I went into unclear. Not only does this make the program more brainlike, but it also makes sure one node doesn’t become way too important while another node becomes completely unclear. The heavy malignant weighting again is just to make sure cancer patients are diagnosed correctly because those are the diagnostics that are going to save lives. And then, something that’s really cool about my program is the inconclusive logic. Often in computer science having no answer is equivalent to failure, but for medicine that’s certainly not the case.
So I researched and I tried to find a good way to implement the inconclusive logic. What I found is that a few programs that are doing inconclusive analysis actually determine all of the you’ll the sigmoid curve that’s right next to the inconclusive logic all of those values that are in the middle part would be deemed inconclusive. And you saw the fine needle aspirates They are difficult to diagnose. So a lot of those values actually fall within that middle part. So if I had done my inconclusive logic that way,.
It wouldnx27t be a usable tool. So instead I simultaneously create 10 different neural networks. Since all the neural networks learn a little bit differently just like we all learn a little bit differently for the truly inconclusive samples they get different results. So they take a vote. If they don’t all agree, then that’s how a mass is deemed inconclusive. I talked a little bit about commercial products. What I mean by that is there are commercial software packages that allow you to create a neural network without having to code it.
So it’s kind of like Excel. You can go in, create a graph and you can pick the colors in the scale, but you don’t have to code it and draw a graph. You can do that with neural networks as well. You can pick the learning rate and the training technique. But of course you are limited a little bit on the customization factors. So I tested three of the leading commercial brands. And then I also created my own network in Java. And I found my own network to be about 5 more sensitive to malignancy.
Than the best of those commercial products. I would like to run through my application. This is what you find if you go on Cloud4Cancer.appspot. It’s in GoogleApp Engine and you are free to visit it if you would like. So the way it works is doctors get these dropdown menus. And they pick on a scale of 1 to 10. So let’s say, the doctor is going to pick the mass. So let’s say, they think it’s a 4. This means that the attribute is not clearly benign. It’s supporting benign unclear, but not definitive.
Doctors would then pick that and go through and do that for all of the corresponding inputs, and almost instantaneously they receive response. In fact this demonstration is a little bit slower than what you would find online. The network is working really well. It’s 99.1 sensitive to malignancy. This is a number I keep coming back to you because it’s huge. This is a number that means it could save lives and a number that means it may be hospitalready with more data and more tests. I still retain a 96.63 specificity to benign masses.
And this is also important because this means it’s still diagnosing the majority of patients who don’t have cancer correctly. So I’d like to take a minute to tell you about my future plans. Ever since winning the Google Science Fair I’ve had this amazing platform that I can use to share my research with the rest of the world. So I’m actually getting this beta tested in hospitals. Linking our medical center up in Philadelphia is giving me more samples because as the neural network gets more samples it gets smarter.
Because it has more experiences to learn from. An institute in Italy is also testing my network against 400 dubious samples. If it does well this could prove to be a partnership where I get up to 15,000 more samples to add to my network. I am also working on extending this program to other types of medical diagnostics. I recently acquired some new ovarian cancer data. So I am trying to make the neural network extend to that because I finally got this platform that’s working really well. So I’m really excited about the future of my program.