CAN WE DIAGNOSE GLAUCOMA WITHOUT DOMAIN KNOWLEDGE?
Not until machines become sentient.
CAN WE DIAGNOSE GLAUCOMA WITH DOMAIN KNOWLEDGE?
We can try...
- Obtain Labled data of positive and negative glaucoma diagnosis images
- Split into train and test sets
- Preprocess Images
- Flatten, greyscale, downsample
- Train network
- Test on test dataset
- Analyze results
IMPORT PACKAGES
In [3]:
%matplotlib inline
import matplotlib
import numpy as np
import matplotlib.pyplot as plt
import scipy
import scipy.stats
import random
import math
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation
from keras.optimizers import SGD
from keras.utils.visualize_util import plot
from tabulate import tabulate
from scipy import ndimage
matplotlib.rcParams['figure.figsize'] = (20.0, 10.0)
BUILD NETWORK
In [5]:
def FFN(inputDim, outputDim):
ffn = Sequential()
ffn.add(Dense(1024, input_dim=inputDim, init='uniform', activation='tanh'))
ffn.add(Dense(1024, input_dim=1024, init='uniform', activation='tanh'))
ffn.add(Dense(outputDim, input_dim=1024, init='uniform'))
ffn.add(Activation('sigmoid'))
return ffn
In [6]:
def downSample(flatPicture,factor):
newImage = np.zeros(len(flatPicture)/factor)
i = 0;
for iPixel in range(len(flatPicture)):
if (iPixel%factor==0) & (i<len(newImage)):
newImage[i] = flatPicture[iPixel]
i+=1
return newImage
In [7]:
def TrainNetwork(compiledNetwork, trainX, trainY, epochs):
for iEpoch in range(epochs):
compiledNetwork.train_on_batch(trainX, trainY)
if iEpoch%100==0:
print iEpoch
return compiledNetwork
LOAD DATA
- Import
- Greyscale
- Flatten (2D > 1D)
- Scale (min, max c [0,1])
In [9]:
originalGlaucomaTrain = [None]*10
trainX = np.zeros((20,8185))
for i in range(10):
if i<9:
iPic = '0' + str(i+1);
else:
iPic = '' + str(i+1);
face = ndimage.imread('/Users/kaftand/Downloads/glaucoma/' + iPic +'_g.JPG',flatten=True)
print '/Users/kaftand/Downloads/glaucoma/' + iPic +'_g.JPG'
originalGlaucomaTrain[i] = face;
face = face.flatten()
trainX[i,:] = downSample(face,1000)/256.0
In [10]:
originalHealthyTrain = [None]*10
for i in range(10):
if i<9:
iPic = '0' + str(i+1);
else:
iPic = '' + str(i+1);
face = ndimage.imread('/Users/kaftand/Downloads/healthy/' + iPic +'_h.JPG',flatten=True)
print '/Users/kaftand/Downloads/healthy/' + iPic +'_h.JPG'
originalHealthyTrain[i] = face;
face = face.flatten()
trainX[i+10,:] = downSample(face,1000)/256.0
In [11]:
trainY = np.zeros((20,1))
for i in range(10):
trainY[i] =1;
CREATE AND TRAIN NETWORK
In [12]:
net = FFN(trainX.shape[1],1)
In [13]:
net.compile(loss='binary_crossentropy', optimizer="SGD");
In [14]:
trainedNet = TrainNetwork(net, trainX, trainY, 2000)
In [15]:
testX = np.zeros((10,8185))
originalGlaucomaTest = [None]*5
originalHealthyTest = [None]*5
for i in range(5):
if i<9:
iPic = '' + str(i+11);
else:
iPic = '' + str(i+11);
face = ndimage.imread('/Users/kaftand/Downloads/glaucoma/' + iPic +'_g.JPG',flatten=True)
print '/Users/kaftand/Downloads/glaucoma/' + iPic +'_g.JPG'
originalGlaucomaTest[i] = face
face = face.flatten()
testX[i,:] = downSample(face,1000)/256.0
for i in range(5):
if i<9:
iPic = '' + str(i+11);
else:
iPic = '' + str(i+11);
face = ndimage.imread('/Users/kaftand/Downloads/healthy/' + iPic +'_h.JPG',flatten=True)
originalHealthyTest[i] = face;
face = face.flatten()
testX[i+5,:] = downSample(face,1000)/256.0
print '/Users/kaftand/Downloads/healthy/' + iPic +'_h.JPG'
VISUALIZE RAW TRAINING DATA
In [16]:
for iPic in range(10):
f, axarr = plt.subplots(1, 2)
axarr[0].imshow(originalGlaucomaTrain[iPic],cmap='Greys_r')
axarr[0].set_title('Glaucoma Train')
axarr[0].get_xaxis().set_visible(False)
axarr[0].get_yaxis().set_visible(False)
axarr[1].imshow(originalHealthyTrain[iPic],cmap='Greys_r')
axarr[1].set_title('Healthy Train')
axarr[1].get_xaxis().set_visible(False)
axarr[1].get_yaxis().set_visible(False)
VISUALIZE RAW TEST DATA
In [17]:
out = trainedNet.predict(testX)
correctTestY = [True,True,True,True,True,False,False,False,False,False]
for iPic in range(5):
f, axarr = plt.subplots(1, 2)
axarr[0].imshow(originalGlaucomaTest[iPic],cmap='Greys_r')
axarr[0].set_title('Glaucoma Test ID ' + str(out[iPic]>.5))
axarr[0].get_xaxis().set_visible(False)
axarr[0].get_yaxis().set_visible(False)
axarr[1].imshow(originalHealthyTest[iPic],cmap='Greys_r')
axarr[1].set_title('Healthy Test ID ' + str(out[iPic+5]>.5))
axarr[1].get_xaxis().set_visible(False)
axarr[1].get_yaxis().set_visible(False)
ANALYZE RESULTS
In [18]:
def identification(pred, act):
TP=0;
TN=0;
FP=0;
FN=0;
for i in range(len(act)):
if act[i]:
if pred[i]:
TP+=1.0;
else:
FN+=1.0;
else:
if pred[i]:
FP+=1.0;
else:
TN+=1.0;
return TP,TN,FP,FN
def sensitivity(TP, FP):
return TP/(TP+FN)
def specificity(TN, FP):
return TN/(TN+FP)
def accuracy(TP, TN, FP, FN):
return (TN + TP)/(TP + TN + FP + FN)
In [19]:
idTable = np.asarray([range(10),correctTestY, (out>.5).flatten()]).T.tolist()
print tabulate(idTable, headers = ['Picture #','Glaucoma Actual', 'Glaucoma Predicted'])
TP,TN,FP,FN = identification((out>.5).flatten(), correctTestY)
sens = sensitivity(TP, FP)
spec = specificity(TN,FP)
acc = accuracy(TP,TN,FP,FN)
specTable = {"Sensitivity": [sens],"Specificity": [spec], "Accuracy" : [acc]}
print ''
print tabulate(specTable,headers = 'keys')
NEAT! WE DID IT!
In [ ]: