Hello! Below you can find a outline of how to reproduce my solution for the RSNA Pneumonia Detection Challenge competition. If you run into any trouble with the setup/code or have any questions please contact me at jiyuanfeng@imsightmed.com

---

ARCHIVE CONTENTS

• classfication_module: classify each image(Normal, No lung, Lung Opacity) ,the folder contain the train.py and test.py. and the Usage will recommend in the readme.md in folder.
• detection_module1: 2 stage detecter, the code is borrowed from (https://github.com/matterport/Mask_RCNN), also, the Usage will recommend in its README.md file.
• detection_module2: 2 stage detecter, the code is borrowed from(https://github.com/tensorpack/tensorpack/tree/master/examples/FasterRCNN), also, the Usage will recommend in its README.md file.
• detection_module3: 2 stage detecter, the code is borrowed from(https://github.com/roytseng-tw/Detectron.pytorch), also, the Usage will recommend in its README.md file.
• emsemble_module: in this module, we use the classifacation results to filter the detection results. the usage will recommend in the sub README.md

---

HARDWARE:

• Intel(R) Xeon(R) CPU E5-2697 v4 36 cores 8 Nvidia Titan Xp
• Ubuntu 16.04
• necessary 3rd-party software: the repository below can be installed by pip
  • classfication_module:
    • tensorboard==1.9.0
    • tensorflow-gpu==1.9.0
  • detection_module1:
    • tensorflow>=1.3.0
    • keras>=2.0.8
    • opencv
    • pydicom
    • tqdm
    • imgaug
  • detection_module2:
    • python3
    • opencv
    • tensorflow >=1.6
    • pycocotools
    • tqdm
  • detection_module3:
    • pytorch>=0.3.1
    • torchvision>=0.2.0
    • cython
    • matplotlib
    • opencv
    • pyyaml
    • packaging
    • pycocotools
    • tensorboardX
  • the more detail requirement.txt is under each module folder

---

How to train our model and make predictions

We have four modules together, one for classification, the other three for detection.

Classification model is trained as usual with 3 different classes(normal, lung opacity and not normal no opacity). Detection models are first fine-tuned from imagenet weights with positive images, than again with both positive and negative images.

• For classfication module:

1.prepare csv for training and testing as the example file in train_csv and test_csv ( csv file is recognised by folder, unnecessary files need to be deleted from the fold) 2.Change the directory in train.py to corresponding folder name for training and run train.py. The result model will be produced in ./result/classification_model/ 3. Change the directory in predict.py to corresponding folder name for testing and run test.py. Prediction result will locate in ./result/test_result/ 4. And more details are describe in the README.md in this module 's folder

• For detection_module1:

1. To train: cd detection_module1/RSNA python train.py
2. To Predict: cd detection_module1/RSNA python predict.py
3. And more details are describe in the README.md in this module 's folder

• For detection_module2:

1. To train: python train.py --config
   DATA.BASEDIR=/path/to/data_prepare/
   DATA.IMGDIR=images
   --load imageNet_model/ImageNet-ResNet101.npz Options can be changed by either the command line or the config.py file.
2. To predict: To predict on image dir: python train.py --config TEST.RESULT_SCORE_THRESH=0.5 TEST.FRCNN_NMS_THRESH=0.1
   --load /path/to/model.index --predictDir path/to/IMAGE_TEST_DIR Then, it will generate a csv file(./results/test_res.csv).
3. And more details are describe in the README.md in this module 's folder

• For detection_module3:

1. To train: python ./tools/train_net_step_rsna.py
   --dataset=rsna2018_pos_neg
   --cfg ./configs/RSNA/e2e_frcnn.yaml
   --use_tfboard
   --load_ckpt /path/to/model
2. To predict: python ./tools/create_submition.py
   --dataset rsna2018
   --submit_threshold 0.4
   --cfg ./configs/RSNA/e2e_frcnn.yaml
   --load_ckpt /path/to/model
   --image_dir path/to/IMAGE_TEST_DIR Then, it will generate a csv file(./Submission/submission.csv).
3. And more details are describe in the README.md in this module 's folder

• For emsemble_module:

1. to filter false positive samples python filter.py
   -predsfile=solution/result_1.csv
   -classfile=solution/test3000_v1p1.csv
   -threshold=0.55
2. to ensemble results python fusion.py -model_num=11
   -iou_threshold=0.3 -vote_threshold=7
3. the details are describe in the README.md in this module 's folder