Pooja Kannappan, ninth-grade student at BHHS, competed at the 2019 Science and Engineering Fair of Metro Detroit (SEFMD), winning a First Place Award. Kannappan's project "Detecting Metastatic Cancer in Digital Pathology Scans Using Machine Learning" also received an Intel Award Certificate for outstanding achievement in the category of Computer Science.
Kannappan said interest in the project came from a desire to make the current cancer diagnostic process better. "Delay of diagnosis and diagnostic errors are some of the major factors in cancer mortality rates. Histopathological analysis of whole-slide images is one of the most widely used techniques for early diagnosis of cancers," Kannappan explained. "Right now, the cancer diagnosis is done by a pathologist sitting in a lab and the analysis done by these pathologists is very specialized and time consuming and causes fatigue. This increases the chances of diagnostic error, so I wanted to be able to use technology to make it easier.." . The project used a machine learning algorithm to automate the task of detecting metastatic cancer from histopathological images more accurately and efficiently. Kannappan explained the process took several resources to achieve the desired outcome. "I had to train my model on a set of images. One of the biggest limitations is not having large dataset for training, so you have to use open source material which in my case was lymph node metastasis data as that is what my model is trained on and that is what it can identify."
The process was relatively complex. "I used the histopathologic cancer detection dataset from Kaggle," said Kannappan. To achieve a high prediction accuracy, Convolutional Neural Networks (CNN) need to be trained on a large dataset which can increase training time significantly. In order to speed up the training process and improve the accuracy in detection, I utilized transfer learning. I trained/experimented with various state of the art CNNs such as InceptionV3, ResNet50, Xception, and NasNetMobile, to create my machine learning models. I used Keras API with TensorFlow backend to create my model. To execute my machine learning experiments efficiently, I used Google Colab Notebook with Graphic Processing Units, or GPU, Python 3 runtime."
In the end, the methodology was a success. By comparing the AUC value of each model, Kannappan was able to conclude that the transfer learning model using NasNetMobile + Xception delivered the highest accuracy in metastatic cancer detection with about 96% accuracy.
This is not Kannappan's first competition. "I've been doing science projects since elementary school, but I started competing in SEFMD in 6th grade," Kannappan said. All of the projects have had to do with the intersection of science and technology. "I made a remote health heart monitoring system using the internet of things that allows someone in a rural area to monitor their vital signs and send it to doctors and their family. I also built an artificial intelligence in health care app that I coded in SWIFT, which is an iOS application. It allows people in rural areas with little to no access to doctors to upload an image to an app to tell if they if they have diabetic retinopathy, which affects a lot of people in poorer countries and causes blindness."
The SEFMD is not a typical science fair. Entrants do need a trifold board and presentation, but the judges are professionals in the category that they are evaluating. "There were a few other projects that used machine learning. At least five to six judges knew a lot about machine learning and artificial intelligence as that is their career path and what they are currently working on. Those kinds of judges look through your code and pick out something like asking about your parameters. They want to make sure that you understand what you did even if you used resources. There were other judges who asked about your interests, like whether or not you were interested in cancer and what your thoughts were on our current diagnosis process, how you got inspired to do this project, maybe if you have read a particular book. Those judges want to get to know you as a person."
The SEFMD runs from about 8 a.m. to 5 p.m. with two parts. In the morning, there are seven to eight judges who ask specific questions for about 10 minutes each. After lunch, everyone looks to see if they have received a pink paper designating them as a finalist. Then, finalists have to be there for the afternoon session during which they meet different judges.
The event is very popular, Kannappan said. "There are a lot of people! Everyone who comes to this fair is really interested in science and wants to win this. In my presentation, I tried to be extremely honest and talk about the things that didn't go well or turn out exactly how I wanted. They want to see the process and see that this person is genuinely interested in this. Not only does your project have to be impressive and have a lot of background to it, but they want to see where you got this project idea from and how did you get through your problems because that is important to any science and engineering fields."
Kannappan expects to continue entering the competition throughout high school and to take classes at BHHS that will support these passions. Kannappan also wants to turn these experiences into opportunities to give back. "One of my ideas for my MYP project is to reflect my interests by teaching younger kids coding and mentor them. We get introduced to it with the Hour of Code, but there's not really a lot of opportunity between that and the computer science classes you can take at BHHS."
After that, Kannappan doesn't know exactly what the future holds, but knows that the mix of science, technology and business will be part of it. "I want to use technology to help the medical field and incorporate business. I want to have a question in mind and look at how to solve it."