CheXExperts implementation in PyTorch

Multi expert fusion disease diagnosis model CheXExperts achieved an AUC score of 0.85 and an IoR score of 0.75 in the CXR14 dataset.

Prerequisites

• Python 3.7
• matplotlib==3.4.3
• multimethod==1.8
• numpy==1.20.3
• opencv_python==4.5.5.64
• Pillow==9.5.0
• pycocotools==2.0.4
• pyswarms==1.3.0
• PyYAML==6.0
• scikit_image==0.19.3
• scikit_learn==0.24.2
• scipy==1.7.1
• torch==1.10.2
• torchvision==0.11.3
• tqdm==4.62.3
• ttach==0.0.3

Preparation

• Download the ChestX-ray14 database from here
• Unpack archives in separate directories (e.g. images_001.tar.gz into images_001)
• Download the trained models and cropped chest X-ray images here
• Unpack segmentations.tar.gz to the same level directory as ChestX-ray14 database.
• Move the best_auc_model26-0.8419458151267506.pth.tar to the CheXExperts/checkpoints/withGAA folder.
• Move the best_auc_model29-0.8544614168758137.pth.tar to the CheXExperts/checkpoints/withoutGAA folder.
• Move the csv_retinanet_epoch3.pt to the CheXExperts/retinanet/models/trained_without_neg_sample_res101 folder.

Usage

• For verifying CheXExpert: Open the CheXExperts\cfgs\chexnet++.yaml and edit the following fields to your own dataset directory:
  images_path: D:\dataset\CXR14\images
  segment_path: D:\dataset\CXR14\segmentations
  Run python Main.py to run verifying.

• For Training CheXExperts:
  1.Training The CheXMHNet without GAA.
  Copy the CheXExperts\checkpoints\withoutGAA\chexnet++.yaml to CheXExperts\cfgs\ , backup the original chexnet++.yaml file.
  Open the chexnet++.yaml and edit the following fields to your own dataset directory:
  images_path: D:\dataset\CXR14\images
  segment_path: D:\dataset\CXR14\segmentations
  Run python Main.py to run Training

  2.Training The CheXMHNet with GAA.
  Copy the CheXExperts\checkpoints\withGAA\chexnet++.yaml to CheXExperts\cfgs\ , backup the original chexnet++.yaml file.
  Open the chexnet++.yaml and edit the following fields to your own dataset directory:
  images_path: D:\dataset\CXR14\images
  segment_path: D:\dataset\CXR14\segmentations
  Run python Main.py to run Training

Results

Experimental environment

GPU RTX 3060 12GB
CPU I5 10400F
Mem 16GB