Code for Discriminative Subtyping of Lung Cancers from Histopathology Images via Contextual Deep Learning.
We propose to use discriminative subtypes: latent representations of samples which provide optimality crtierion via downstream predictive performance. Because discriminative subtypes are defined by a machine learning criterion of downstream performance, these latent representations can be optimized via end-to-end training and then used in diverse downstream tasks.
This repository contains code for running Contextualized Explanation Networks (CEN) on TCGA/TCIA. Generic CEN code is available at https://github.com/alshedivat/cen.
Available models include:
To facilitate experiments, training and eval scripts are included.
Data handlers are provided for TCGA and several other datasets.