an iScience's paper (read online):

@article{PREECHAKUL2022103933,
title = {Improved image classification explainability with high-accuracy heatmaps},
journal = {iScience},
volume = {25},
number = {3},
pages = {103933},
year = {2022},
issn = {2589-0042},
doi = {https://doi.org/10.1016/j.isci.2022.103933},
url = {https://www.sciencedirect.com/science/article/pii/S2589004222002036},
author = {Konpat Preechakul and Sira Sriswasdi and Boonserm Kijsirikul and Ekapol Chuangsuwanich},
keywords = {Artificial intelligence, Computer science, Signal processing}

CAM methods are used to produce heatmaps for explaining a deep classifier. However, these heatmaps come in low resolution hindering explainaibilty of the classifier.

PYLON extends the deep classifier allowing CAM methods to generate much higher resolution heatmaps resulting in much more accurate explanation.

High resolution heatmaps:

PYLON's architecture:

• Pytorch implementation of PYLON (see pylon.py)
  • Pretrained weights
• Additional results and heatmaps
• Code to reproduce the main results

1. Picked localization images (256 x 256) (~20)

   • PYLON
   • Backbone
   • DeeplabV3+
   • DeeplabV3+ (No GAP)
   • FPN (BN)
   • Li 2018
   • PAN
   • UNET

2. Picked localization images (512 x 512) (~20)

   • PYLON
   • Backbone
   • FPN (BN)
   • Li 2018

3. All localization images (256 x 256) (~1000 images)

It was tested with Pytorch 1.7.1.

conda create -n pylon python=3.7
conda activate pylon
conda install pytorch=1.7.1 torchvision cudatoolkit=11.0 -c pytorch

Install other related libraries:

pip install -r requirements.txt

You need to download the Chest X-Ray 14 dataset by yourself from https://nihcc.app.box.com/v/ChestXray-NIHCC.

Extract all the images into a single big directory data/nih14/images, containing 100k images.

Download the DICOM version from Kaggle https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/overview

You need to convert them into PNG. The script is provided in scripts/convert_dcm_to_png.py. The conversion will not alter the aspect ratio it will aim for maximum 1024 either width or height.

Put all the PNG files into directory data/vin/images. The final directory structures:

Download the dataset from http://host.robots.ox.ac.uk/pascal/VOC/voc2012/. (Only train & val datasets)

Extract it to data/voc2012. You should see the following directory structures:

data/voc2012
- VOCdevkit

The main run files are train_nih_run.py (see train_nih.py for reference), train_vin_run.py (see train_vin.py for reference), and train_voc_run.py (see train_voc.py for reference). The files describe experiments to be run. They are straightforward to edit. Make changes or read before you run:

python train_nih_run.py

python train_vin_run.py

python train_voc_run.py

In order to compare with a wide-range of CAM methods, run the following scripts:

python eval_cam.py

The results will be shown at eval_loc/<cam_mode>/<model_name>

Stats will be available at eval_auc and eval_loc.

The figures will be available at figs/picked and figs/all.

You can change the config in mlkitenv.json:

{
    # available GPUs 
    "cuda": [0],
    "num_workers": 8,
    # number of parallel jobs
    "global_lock": 1
}