RSNA Intracranial Hemorrhage Detection

• This is the project for RSNA Intracranial Hemorrhage Detection hosted on Kaggle in 2019.
• It finished at 3rd place in the competition.

Experiments

exp	window	input	output	comment
exp7_seres_precenter_512	subdural	st	st	for external data pseudo labeling
exp10_seres_threecenter_rescale_512	3types	st	st
exp16_seres_concat	subdural	st-1, st, st+1	st
exp17_7_external	subdural	st	st	with external data
exp18_seres_concat_three	3types	st-1, st, st+1	st
exp19_seres_doubleconcat	subdural	st-2, st, st+2	st
exp21_seres_doublepre	subdural	st-2, st-1, st	st
exp22_seres_doublepost	subdural	st, st+1, st+2	st
exp23_seres_doublepre_three	3types	st-2, st-1, st	st
exp24_seres_doublepost_three	3types	st, st+1, st+2	st
exp25_seres_conc3	subdural	mean(st-3, st-2, st-1), st, mean(st+1, st+2, st+3)	st
exp26_seres_concall_prepost	subdural	mean(all st's), st, mean(st-1, st+1)	st
exp27_seres_conc5	subdural	mean(st-5, st-4, st-3, st-2, st-1), st, mean(st+1, st+2, st+3, st+4, st+5)	st
exp28_seres101_16	subdural	st-1, st, st+1	st	seresnext101
exp32_seres_conc_any	subdural	st-1, st, st+1	st	predict 5 classes and fill_“any” on max prediction.
exp34_seres_threetarget	subdural	st-1, st, st+1	st-1, st, st+1
exp36_seres_double_threetarget	subdural	st-2, st, st+2	st-2, st, st+2

Usage

1. Download datasets in input dir.

2. preprocessing

sh bin/preprocess.sh

3. train

sh bin/train.sh

4. predict

sh bin/predict.sh

4. stacking sry this part of code is super dirty...

run kernel.
then download output file in output dir.

sh bin/stack.sh

Usage Demo

only predition with pretrained models. This pipeline use user stacking only. The private score=0.04393.

1.Download models from kaggle datasets.

2.Run prediction

sh bin/demo.sh

Hardware and environment

GCP deep learning vm

• gpu: V100*2
• cpu: n1-standard-16(16 vCPU, 60GM RAM)

Install

pip install -r requirements.txt

License

The license is MIT.