Online Onychomycosis Model - RESTful API
Offline Onychomycosis Model - BVLC caffemodel

- Python (version 2.x is better than 3.x)
- Internet access

- api.zip


1) Install Python 2.7
Download Anaconda2 64bit 2.7 (https://www.continuum.io/downloads#windows) and install.
Set installation path to "c:\anaconda2" or "c:\python27"
Be sure to check "Add anaconda to my PATH environment variable" during installation

2) Download api.zip and extract api.zip

3) Resolve dependencies (requests and matplotlib)
c:\api\> pip install --upgrade pip
c:\api\> pip install requests matplotlib
or
c:\api\> setup.bat

4) Run Python scripts (api.py, roc.py)
c:\api\> python api.py
c:\api\> python roc.py

# Source code - api.py

###
### Nail model API
### Han Seung Seog (whria78@gmail.com) / http://whria.net / http://medicalphoto.org
### 2017-3-5
###

import requests
import pickle
import os
import sys

# for python2 python3 compatibility
try:
    import urllib.parse as myurl
except ImportError:
    import urllib as myurl
from io import BytesIO

# for unicode support
def custom_filename(x):return myurl.quote(x)

def modelnail(url,username,image_path):
    url=url+'/api'
    data_={'user':username}
    file_={'file':(custom_filename(os.path.basename(image_path)),open(image_path,'rb').read())}

    res=requests.post(url,data=data_,files=file_)
    f=BytesIO(res.content)

    # for python2 python3 compatibility
    if sys.version_info[0] >= 3:
        result=pickle.load(f,encoding='bytes')
    else:
        result=pickle.load(f)
    return result

def get_deploy(url):
    url=url+'/deploy'
    res=requests.get(url)
    f=BytesIO(res.content)
    result=pickle.load(f)
    return result

### main server ### 
server_url='http://nail.medicalphoto.org:8001'
### backup server ### 
#server_url='http://backup.medicalphoto.org:8001'

username='HAN'

### Get list of indexes ###
list_deploy=get_deploy(server_url)

### Run Model ###

test_filename='test.jpg'
if len(sys.argv)>1: 
    test_filename=str(sys.argv[1])
if (os.path.exists(test_filename)==False):
    print (test_filename+' is not exist')
    sys.exit(0)
   

result=modelnail(server_url,username,test_filename)
success=result[0]
model_result=result[1]

if (success!=True):
    print (model_result)
    sys.exit(0)

for k in range(0,len(list_deploy)):
    print ("\n")
    print ("## Model %d ##" % (k+1))

    for i,deploy_ in enumerate(list_deploy[k]):
        print ("%s : %.4f" % (list_deploy[k][i],model_result[k][i]))

Example 1 - test.jpg

~$ python api.py
or
~$ python3 api.py

## Model 1 ## (ResNet-152)
onychomycosis : 0.0029
normalnail : 0.0016
naildystrophy : 0.9858
melanonychia : 0.0001
onycholysis : 0.0033
others : 0.0063


## Model 2 ## (VGG-19)
onychomycosis : 0.2237
normalnail : 0.0021
naildystrophy : 0.7443
melanonychia : 0.0004
onycholysis : 0.0228
others : 0.0067


## Model 3 ## (Fine Image Selector)
nail : 0.9871  (adequate value of fine image selector)
focus : 0.0126
dumy : 0.0003
skinphoto : 0.0000
index : 0.0000
imagenet : 0.0000


## Model 4 ## (ResNet-152 + VGG-19 + Feedforward network)
onychomycosis : 0.1039
non_onychomycosis : 0.8961

Example 2 - Draw 1) ROC curve (test dataset = B1+C; training dataset = A1) 2) Threshold / Sensitivity / Specificity graph

We created ROC curves, using MatPlotLib (https://matplotlib.org). It will takes about 15 minutes to complete the tasks depending on network bandwidths.

# ResNet-152 alone ( ~$ python roc.py 1 )

# ResNet-152 + VGG-19; arithmetic mean ( ~$ python roc.py 0 ; ~$ python roc.py )

# ResNet-152 + VGG-19; feedforward neural networks ( ~$ python roc.py 2 )

~$ python roc.py
or
~$ python3 roc.py

DX LIST :

...

/home/ai/edrive/http/medicalphoto/nail/B1C dataset/C/onychomycosis/C_OM (51).JPG
/home/ai/edrive/http/medicalphoto/nail/B1C dataset/C/onychomycosis/C_OM (31).JPG
/home/ai/edrive/http/medicalphoto/nail/B1C dataset/C/onychomycosis/C_OM (32).JPG

AUC : 94.86, MAX J score : 78.62
MAX sen / spe : 88.79 / 89.83 Threshold : 0.3263

- Python
- BVLC PyCaffe

- Figshare.com (3.7 Gigabytes)


1. Download the zip file. It contains,

1) Model Onychomycosis; Models trained with A1, A2, and E dataset + the feedforward neural networks model
2) Python example and batch files that generate ROC curves
3) Prebuilt 64bit windows pycaffe - "caffe" folder

2. Install BVLC PyCaffe
Tutorial : Install PyCaffe

3. Run examples

Fig3C-A1-ensemble.bat
Fig3C-A1-ResNet.bat
...
...
...
Fig4C-A2-ResNet.bat
Fig4C-E-ResNet.bat