/CheXplain-IBA

MICCAI 2021 | Explaining COVID-19 and Thoracic Pathology Model Predictions by Identifying Informative Input Features

Primary LanguageJupyter NotebookMIT LicenseMIT

Explaining COVID-19 and Thoracic Pathology Model Predictions by Identifying Informative Input Features

This is the source code for our paper. The repository is divided into 3 parts, IBA contains code for attribution methods (Information Bottleneck), model contains training script; dataset (NIH ChestXray, BrixIA) and learned model; evaluation contains codes for quantitative evaluations.

Setup

  1. Make sure you have conda installed, then create an environment using conda env create -f environment.yml.
  2. Follow the installation guide from IBA to install Information Bottleneck method.
  3. Download BrixIA and NIH ChestXray images, the labels is included in this repository in model/labels

Usage

We provide several Jupyter notebooks in each sub-folder (IBA/notebooks, model/notebooks)

Model Training

In model/results we provide trained model for NIH ChestXray, BrixIA regression, and BrixIA classification scheme In model/notebooks their are notebooks to train, fine tune, and evaluate models

Model Attribution

We have included various notebooks to run Information Bottleneck on ImageNet, NIH ChestXray, and BrixIA dataset. The notebooks can be found in IBA/notebooks

Evaluate Attribution Maps

To evaluate the correctness of attribution maps, we provide two quantitative evaluations, which are sensitivity-N and insertion/deletion. To batch run evaluations, use source run_insertion_deletion.h and source run_sensitivity_n.h respectively. Before run the evaluations, make sure the variable defined inside bash scripts have correct path assigned