Short description

Breast cancer diagnosis using machine learning

July 2016

Think it.

"Breast cancer affects 1 in 8 women. According to a University of Washington study, pathologists offer conflicting diagnoses for the same histological data 25% of the time. Less invasive cancer is even more difficult to identify, resulting in an abundance of false positives. Such diagnoses can lead to women taking unnecessary steps such as radiation therapy or surgery. Considering this, we developed a diagnostic tool for pathologists called Onkos, to give them greater confidence in their own predictions.

Website I built for the platform

Build it.

Our model uses Convolutional Neural Networks (CNNs) to learn from histological images on whether they indicate breast malignancy or not. CNNs consist of a sequence of convolutional layers. The network starts with simple object features such as edges and as it moves to higher convolution layers, it is able to describe more complex features such as shapes and entire object. For this reason, CNN is well-suited to image classification. We developed our model using Tensorflow framework and trained the images on a NVIDIA GPU Instance on Amazon AWS.

Photo: Ben Auerbach, Winnie Cheng, Tapomay Dey, Ben Issenmann, Chitrang Talaviya (not in photo). We look pretty good for coding all weekend without sleep!

Link to original post by Winnie Cheng here.

Lessons learned:

  • Hackathons are amazing learning experiences. Pressure bonds teammates and pushes you to exceed yourself.
  • When the roles are clear inside a team (I worked on product and design), people are more effective
  • Hackathons are about building cool stuff, not necessarily useful stuff. This is where you can apply the latest technology you've learned.