Deep Learning and Cancer Research | Two Minute Papers #64

Dear fellow scholars, this is two minute papers with let's try to assess the workflow of this piece of work in the shortest possible form. The input is images of cells and the output of the algorithm is a decision that tells us which one of these are cancer cells. As the pipeline of the entire experiment is quite elaborate, we'll confine ourselves to discuss the deep learning related step at the very end. Techniques prior to this one involved adding chemicals to blood samples. The problem is that these techniques were not so reliable and that they also destroyed the cells, so it was not possible to check the samples later. As the title of the paper says, it is a label free technique. Therefore, it can recognize cancer cells without any intrusive changes to the samples. The analysis happens by simply looking at them. To even have a chance at saying anything about these cells, domain experts have designed a number of features that an educated decision. For instance, they like to look at refractive indices that tell us how much light slows down when passing through cells. Light absorption and scattering properties are also recognized by the algorithm. Morphological features are also quite important as they describe the shape of the cells and they are among the most useful features for the detection procedure. So the input is an image, then come the high level features and the neural networks help locating the cancer cells by learning the REL exactly what values for these high level features lead to cancer cells. The proposed technique is significantly more accurate and consistent in the detection than previous techniques. It is of utmost importance that we are able to do something like this on a mass scale because the probability of curing cancer depends greatly on which phase we can identify it. One of the most important factors is early detection. And this is exactly how deep learning can aid us. To demonstrate how important early detection is, have a look at this chart of the ovarian cancer survival rates as a function of how early the detection takes place. I think the numbers speak for themselves, but let's bluntly state the obvious. It goes from almost surely surviving to almost surely dying. By the way, they were using L2 regularization to prevent overfitting in the network. We have talked about what each of these terms mean in a previous episode. I've put a link for that in the description box. 95% success rate with a throughput of millions of cells per second. Wow. Bravo. A real two minute paper style hat tip for the authors of the paper. It is really amazing to see different people from so many areas working together to defeat this terrible disease. Engineers create instruments to be able to analyze blood samples. Doctors choose the most important features and computer scientists try to find out the relation between the features and illnesses. Great strides have been made in the last few years and I am super happy to see that even if you are not a doctor and you haven't studied medicine, you can still help in this process. That's quite amazing. A big shout out to Cram, who has been watching two minute papers since the very first episodes and his presence has always been ample with insightful comments. Thanks for being around and also thanks for watching and for your generous support and I'll see you next time.