Cancer Diagnosis via AI: sparse vs. deep learning

Team: 1005

School: Los Alamos High

Area of Science: Computational Medicine


Problem Definition:

Detecting Cancer using a neural network, trained on pathological tissue slides had been done before via deep learning neural networks. These deep neural networks have been shown to be less stable to pixel-level-changes than sparse coding neural networks. The pixel-level-changes, which cause misclassification, are known as “adversarial examples”. In a previous publication, I helped show that sparse coding is more robust to transferable adversarial examples than traditional deep learning. When classifying tissue as tumorous or normal, stability is a must-have, thus exposing the need to develop sparse coding method for cancer detection. Sparse coding uses sparsity to achieve under-completeness, rather than forcing the model to use dimensional reduction only. I will compare deep learning models and sparse coding models on the problem of tumor detection.

Progress to Date:

Since beginning this project, I have written multiple command line tools which train different kinds of neural networks. A major challenge here is simply dealing, efficiently, with the massive size of histological images (400 GB compressed). Previously a co-worker sparse coded the dataset for me, producing thousands of compressed versions of these tiles. The problem I am solving for this supercomputing challenge project is to predict the classification on each pixel as cancer or not-cancer and compare models that do this. For comparison, I created a deep learning autoencoder which returned similar compressions at the latent layer as the sparse coded set. Next, I wrote a classification neural network. I have applied this to both the sparse coded images and then deep-learning latent images. My classifier took in compressed tiles, outputting an image of predictions from 0-1 for each slide (aka a heat map of possible tumorous areas). Both classifiers returned satisfactory diagnosis on select testing tissues. Currently, I am writing the tool to compute ROC AUC, and PR AUC on the outputted predictions from both classifiers. With these metrics, I will compare my models to each other, and the models of other people. Then I will use these metrics to test their robustness under adversarial examples.

Expected Results:

I want to see my models do as well, if not better than the already published ROC AUC’s of other models. I also expect to find the sparse coding model to be more robust to adversarial noise than the deep neural network model.

Mentor: Dr. Garrett T. Kenyon

Team Members:

  Charles Strauss

Sponsoring Teacher: NA

Mail the entire Team