RSNA Intracranial Hemorrhage Detection

This is the source code for the first place solution to the RSNA2019 Intracranial Hemorrhage Detection Challenge.

Journal Link | Cite

Solution write up: Link.

Solutuoin Overview

Dependencies

• opencv-python==3.4.2
• scikit-image==0.14.0
• scikit-learn==0.19.1
• scipy==1.1.0
• torch==1.1.0
• torchvision==0.2.1

CODE

• 2DNet
• 3DNet
• SequenceModel

2D CNN Classifier

Pretrained models

• seresnext101_256*256 [seresnext101]
• densenet169_256*256 [densenet169]
• densenet121_512*512 [densenet121]

Preprocessing

Prepare csv file:

download data.zip: https://drive.google.com/open?id=1buISR_b3HQDU4KeNc_DmvKTYJ1gvj5-3

1. convert dcm to png

python3 prepare_data.py -dcm_path stage_1_train_images -png_path train_png
python3 prepare_data.py -dcm_path stage_1_test_images -png_path train_png
python3 prepare_data.py -dcm_path stage_2_test_images -png_path test_png

2. train

python3 train_model.py -backbone DenseNet121_change_avg -img_size 256 -tbs 256 -vbs 128 -save_path DenseNet121_change_avg_256
python3 train_model.py -backbone DenseNet169_change_avg -img_size 256 -tbs 256 -vbs 128 -save_path DenseNet169_change_avg_256
python3 train_model.py -backbone se_resnext101_32x4d -img_size 256 -tbs 80 -vbs 40 -save_path se_resnext101_32x4d_256

3. predict

python3 predict.py -backbone DenseNet121_change_avg -img_size 256 -tbs 4 -vbs 4 -spth DenseNet121_change_avg_256
python3 predict.py -backbone DenseNet169_change_avg -img_size 256 -tbs 4 -vbs 4 -spth DenseNet169_change_avg_256
python3 predict.py -backbone se_resnext101_32x4d -img_size 256 -tbs 4 -vbs 4 -spth se_resnext101_32x4d_256

After single models training, the oof files will be saved in ./SingleModelOutput(three folders for three pipelines).

After training the sequence model, the final submission will be ./FinalSubmission/final_version/submission_tta.csv

Sequence Models

Sequence Model 1

Sequence Model 2

Path Setup

Set data path in ./setting.py

download

download [csv.zip]

download [feature samples]

Sequence Model Training

CUDA_VISIBLE_DEVICES=0 python main.py

The final submissions are in the folder ../FinalSubmission/version2/submission_tta.csv

Final Submission

Private Leaderboard:

• 0.04383

Reference

If you find our work useful in your research or if you use parts of this code please consider citing our paper:

title = {A Deep Learning Algorithm for Automatic Detection and Classification of Acute Intracranial Hemorrhages in Head CT Scans},
journal = {NeuroImage: Clinical},  
pages = {102785},  
year = {2021}  

TODO

• Pre-trained models
• 2DCNN + SeqModel end-to-end training
• 3DCNN training