RSNA-Pneumonia-Challenge-Models

src

This folder contains the code used for detection and classification

Detection

• To run, navigate to kaggle-rsna/
  • run . create_env.sh to create the environment
• Change paths in run_experiments.sh and config.py to point to output, input csvs, and dicom images

• Assumes input includes csvs with 'name, x, y, height, width, class, Target' info
• Also assumes dicom image input

• Run . run_experiments.sh to train and test model
  • Only runs train and check_metric methods in train_runner.py
  • Edit src/datasets as necessary (only uses detection_dataset, dataset_valid, and test_dateset for train and check_metric)

Classification

• Wildcat
  • Contains wildcat code to run wildcat models
  • Change run_wildcat_m01.sh as necessary to run different models
  • change and run python3 class_grad.py as appropriate to obtain AUC and precision results on the models
    • outputs a csv file with all of the results

models

This folder contains pretrained models, stored in github LFS

Detection

• positive_only_equalized_model
  • model trained on only positive pneumonia cases, where the dicom inputs are histogram equalized
  • performs the best so far (mAP on positive-only test cases is 0.251)
• 3_class_equalize_model
  • model trained on all 3 classes (Normal, Not Normal/Not Pneumonia, Pneumonia), dicoms, histogram-equalized
  • performs poorly (mAP 0.082)

Classification

• Wildcat Models
  • Contains pretrained wildcat models used in the MICCAI paper