a b/README.md 

  1
  
# MURA





  
  

  2
  
MURA(musculoskeletal radiographs) - bone x-ray





  
  

  3
  





  
  

  4
  
Reference: https://stanfordmlgroup.github.io/competitions/mura/





  
  

  5
  





  
  

  6
  
### Prerequite





  
  

  7
  
Python 3.5





  
  

  8
  
<br>TensorFlow 1.8+





  
  

  9
  
<br>keras 2.2.0





  
  

  10
  
<br>numpy 1.14.5





  
  

  11
  
<br>pandas 0.23.3





  
  

  12
  
<br>sklearn 0.19.1





  
  

  13
  





  
  

  14
  
### histogram_equalization usage





  
  

  15
  
keras_preprocessing package 의 image.py 파일에 data augmentation부분이 수정되야함





  
  

  16
  
/home/casper/.local/lib/python3.5/site-packages/keras_preprocessing/image.py 를 해당 image.py로 교체 





  
  

  17
  





  
  

  18
  
```shell





  
  

  19
  
### Clone the repo.





  
  

  20
  
git clone https://github.com/AItrics/MURA.git





  
  

  21
  
cd MURA





  
  

  22
  





  
  

  23
  
### Transform MURA-v1.1 folder to data folder





  
  

  24
  
python3 download_and_convert_mura.py





  
  

  25
  
### training에 들어가는 Input형태로 폴더와 파일을 정라하여 /data폴더에 넣어줌





  
  

  26
  





  
  

  27
  
### Run





  
  

  28
  
python3 train.py





  
  

  29
  





  
  

  30
  
### To evaluate





  
  

  31
  
python3 eval.py





  
  

  32
  
python3 predict.py MURA-v1.1/valid.csv prediction.csv





  
  

  33
  
```





  
  

  34
  
### Performance 





  
  

  35
  
- Ensemble Model





  
  

  36
  
(model1 + model3 + model3  + model5 + model5) /5 로 평균낸 ensemble





  
  

  37
  





  
  

  38
  
|                                  | val_loss  | accuracy  |





  
  

  39
  
| -------------------------------: | :-------- | :---------|





  
  

  40
  
|  DenseNet201(420x420)            | 0.4320    | 0.8332    |





  
  

  41
  
|  DenseNet169(520x520)with cutout | 0.4045    | 0.8313    |





  
  

  42
  
|  InceptionResNetV2(420x420)      | 0.4211    | 0.8341    |





  
  

  43
  
|  DenseNet201(420x420)            | 0.4311    | 0.8177    |





  
  

  44
  
|  DenseNet169(520x520)            | 0.4082    | 0.8307    |





  
  

  45
  





  
  

  46
  
|                 | per image | per study |





  
  

  47
  
| --------------: | :-------- | :---------|





  
  

  48
  
|    accuracy     | 0.831     | 0.857     |





  
  

  49
  
|       f1        | 0.846     | 0.875     |





  
  

  50
  
|    precision    | 0.806     | 0.840     |





  
  

  51
  
|     recall      | 0.889     | 0.914     |





  
  

  52
  
| **cohen_kappa** | **0.661** | **0.708** |





  
  

  53
  





  
  

  54
  





  
  

  55
  
- Single Model (DenseNet169 with Histogram Equalization, batch_size=8,  img_size=420)





  
  

  56
  
@strange://shared/casper/MURA/models/DenseNet169_420_NEW_HIST.hdf5





  
  

  57
  





  
  

  58
  
|                 | single    | 





  
  

  59
  
| --------------: | :-------- | 





  
  

  60
  
|      loss       | 0.4125    | 





  
  

  61
  
|    accuracy     | 0.856     | 





  
  

  62
  
|       f1        | 0.870     | 





  
  

  63
  
|    precision    | 0.866     | 





  
  

  64
  
|     recall      | 0.873     | 





  
  

  65
  
| **cohen_kappa** | **0.705** | 





  
  

  66
  





  
  

  67
  
### data





  
  

  68
  
데이터경로 :





  
  

  69
  
<br>MURA-v1.1 : @strange:/shared/casper/MURA/MURA-v1.1(36808장) 또는 @strange:/shared/casper/MURA/MURA-v1.1.zip 





  
  

  70
  
<br>MURA-v1.0 : @strange:/shared/casper/MURA/MURA-v1.0          또는 @strange:/shared/casper/MURA/mura-v1.0.zip





  
  

  71
  
<br>MURAv1.0 + MURAv1.1(49158장) : @strange:/shared/casper/MURA/data 





  
  

  72
  





  
  

  73
  
### log





  
  

  74
  
로그 보기 예시:





  
  

  75
  
tensorboard --logdir=./logs/DenseNet169_420_NEW_HIST/





  
  

  76
  





  
  

  77
  
### model





  
  

  78
  
모델 경로:





  
  

  79
  
./models