MURA

MURA(musculoskeletal radiographs) - bone x-ray

Reference: https://stanfordmlgroup.github.io/competitions/mura/

Prerequite

Python 3.5

TensorFlow 1.8+

keras 2.2.0

numpy 1.14.5

pandas 0.23.3

sklearn 0.19.1

histogram_equalization usage

keras_preprocessing package 의 image.py 파일에 data augmentation부분이 수정되야함
/home/casper/.local/lib/python3.5/site-packages/keras_preprocessing/image.py 를 해당 image.py로 교체

### Clone the repo.
git clone https://github.com/AItrics/MURA.git
cd MURA

### Transform MURA-v1.1 folder to data folder
python3 download_and_convert_mura.py
### training에 들어가는 Input형태로 폴더와 파일을 정라하여 /data폴더에 넣어줌

### Run
python3 train.py

### To evaluate
python3 eval.py
python3 predict.py MURA-v1.1/valid.csv prediction.csv

Performance

Ensemble Model
(model1 + model3 + model3 + model5 + model5) /5 로 평균낸 ensemble

	val_loss	accuracy
DenseNet201(420x420)	0.4320	0.8332
DenseNet169(520x520)with cutout	0.4045	0.8313
InceptionResNetV2(420x420)	0.4211	0.8341
DenseNet201(420x420)	0.4311	0.8177
DenseNet169(520x520)	0.4082	0.8307

	per image	per study
accuracy	0.831	0.857
f1	0.846	0.875
precision	0.806	0.840
recall	0.889	0.914
cohen_kappa	0.661	0.708

Single Model (DenseNet169 with Histogram Equalization, batch_size=8, img_size=420)
@strange://shared/casper/MURA/models/DenseNet169_420_NEW_HIST.hdf5

	single
loss	0.4125
accuracy	0.856
f1	0.870
precision	0.866
recall	0.873
cohen_kappa	0.705

data

데이터경로 :

MURA-v1.1 : @strange:/shared/casper/MURA/MURA-v1.1(36808장) 또는 @strange:/shared/casper/MURA/MURA-v1.1.zip

MURA-v1.0 : @strange:/shared/casper/MURA/MURA-v1.0 또는 @strange:/shared/casper/MURA/mura-v1.0.zip

MURAv1.0 + MURAv1.1(49158장) : @strange:/shared/casper/MURA/data

log

로그 보기 예시:
tensorboard --logdir=./logs/DenseNet169_420_NEW_HIST/

model

모델 경로:
./models