RSNA-Pneumonia-Detection-Challenge

This repository gathers the code for RSNA Pneumonia Detection Challenge from the Kaggle competition.

We use Faster R-CNN[1], an open competition notebook using Faster R-CNN to implement.

Reproducing Submission

We need to do some pre-preparation for training and testing on the stage 2 dataset.

To reproduce my submission without retrainig, do the following steps:

1. Requirement
2. Repository Structure
3. Inference

Hardware

Ubuntu 18.04.5 LTS

Intel® Core™ i7-3770 CPU @ 3.40GHz × 8

GeForce GTX 1080/PCIe/SSE2

Requirement

If you want to inference without create envs, you can download this notebook and use Colab env. Run the all cells and it will automatically download the final submission.

All requirements should be detailed in environment.yml.

$ cd RSNA-Pneumonia-Detection-Challenge
$ conda env create -f environment.yml
$ pip install ipykernel
$ pip install opencv-python

We use VSCode to open train.ipynb or inference.ipynb, and choose the kernel name "rsnatest". Next, Run all the cells.

Maybe you can use Kaggle or Colab to run these codes.

The jpg images can be downloaded from here and put in root.

The origin .dam can be downloaded from here and put in root.

Repository Structure

The repository structure is:

RSNA-Pneumonia-Detection-Challenge(root)
  +-- input                         # all files need in this task
  |   +-- images                    # training data
  |   +-- samples                   # testing data
  |   +-- stage_2_train_labels.csv  # training labels
  +-- stage_2_test_images           # testing .dcm
  +-- stage_2_train_images          # training .dcm
  +-- models                        # model weights
  +-- calculate_map.py              # ensemble utils
  +-- ensemble.py                   # reproduce my submission file
  +-- inference.ipynb               # for testing 
  +-- train.ipynb                   # for training model
  +-- environment.yml               # yaml file for establishing the environment

Training

To train the model, run train.ipynb:

The "FOLD" parameter is for cross validation. Please modify the number from 0~4 to train.

Trained model will be saved as models/fasterrcnn_resnet50_fpn_pneumonia_detection_best.pth

Inference

Please download these five model weights if you want to reproduce my submission file, and put them in root or the folder you create.

To reproduce my submission file or test the model you trained, run inference.ipynb.

Note that you need to modify MODEL_PATH and the SUBMISSION_PATH for five model weights(SUBMISSION_PATH should not the same for different weights).

Prediction file will be saved as root/{SUBMISSION_PATH}

Finally, you can run the following command to ensemble the five .csv (using conda env "rsnatest")[2].

We also provide the five .csv.

conda activate rsnatest
cd RSNA-Pneumonia-Detection-Challenge

python ensemble.py stage_2_test_images output ensemble.csv 0.csv 1.csv 2.csv 3.csv 4.csv

0.csv 1.csv 2.csv 3.csv 4.csv is the csv you create above step.

Reference

[1] Faster R-CNN notebook

[2] Ensemble Method