I was inspired by http://www.nature.com/nature/journal/v542/n7639/full/nature21056.html to re-train the ResNet v4 model on a prostate cancer MRI data set from http://spiechallenges.cloudapp.net/competitions/7#participate
What it does
Open the app and point your camera to a fullscreen image of an MRI scan of prostate cancer on your computer screen (example images https://imgur.com/a/FCrvt). It will classify the image into one of 5 histological scores, showing the confidence probabilities associated with each of the scores (https://en.wikipedia.org/wiki/Gleason_grading_system)
How I built it
I downloaded the data set and converted the DICOM MRI scans to JPEG files. In python, I trained the Inception tensorflow model with this data set. Since the project was very large, I optimized the algorithm and associated files for mobile.
Challenges I ran into
Training the model with 10000+ images on CPU took very long. There are significant limitations as to the efficacy of the results as I did not try to optimize the hyperparameters.
Accomplishments that I'm proud of
I built this in 20 hours of work since last night and learned a lot the whole way.
What I learned
Training a deep neural network with a large data set requires significant computation power. I learned that this product would work better if I had worked with a team. Since I heard about this competition only a few days ago and started working on it last night, I did not have the input of other minds nor is it a product ready for real medical use.
What's next for Prostate Cancer Detector
Get a team together, rebuild and retrain the model, rigorously test it, consult with doctors and pathologists at UCSF, optimize, get user feedback.