In [ ]:

    

import tensorflow as tf
#import tensorflow.contrib.learn.python.learn as learn
import tflearn 
import scipy as sp
import numpy as np 
import matplotlib.pyplot as plt
import six
from sklearn.metrics import roc_curve
import datetime
%matplotlib inline


    

In [ ]:

    

comData = np.load("./inData/3D-conv/comData.npy")
comClass = np.load("./inData/3D-conv/comClass.npy")


    

In [ ]:

    

def processClassData(classData):
    """
    Process classData.
    
    Returns a one-hot array of shape [len(classData), 2].
    """
    # Convert label data to one-hot array
          
    classDataOH = np.zeros((len(classData),2))
    classDataOH[np.arange(len(classData)), classData] = 1
    
    return classDataOH

comData = comData[..., np.newaxis]
comClassOH = processClassData(comClass)


    

In [ ]:

    

# kfold x-validation...
k = 3
kfoldData = np.array_split(comData, k)
kfoldLabelsOH = np.array_split(comClassOH, k)
kfoldLabels = np.array_split(comClass, k)
print(kfoldData[0].shape)


    

In [ ]:

    

spec = []
sens = []
roc = []


    

In [ ]:

    

for i in np.arange(0,k,1):
    sess = tf.InteractiveSession()
    tf.reset_default_graph()
    tflearn.initializations.normal()

    # Input layer:
    net = tflearn.layers.core.input_data(shape=[None, 50, 19, 17, 1])

    # First layer:
    net = tflearn.layers.conv.conv_3d(net, 8, [5,5,5],  activation="leaky_relu")
    net = tflearn.layers.conv.max_pool_3d(net, 2, strides=2)

    # Second layer:
    net = tflearn.layers.conv.conv_3d(net, 16, [5,5,5], activation="leaky_relu")
    net = tflearn.layers.conv.max_pool_3d(net, 2, strides=2)

    # Fully connected layer
    net = tflearn.layers.core.fully_connected(net, 1024, regularizer="L2", weight_decay=0.001, activation="leaky_relu")
    #net = tflearn.layers.core.fully_connected(net, 1024, regularizer="L2", weight_decay=0.001, activation="leaky_relu")

    # Dropout layer:
    net = tflearn.layers.core.dropout(net, keep_prob=0.5)

    # Output layer:
    net = tflearn.layers.core.fully_connected(net, 2, activation="softmax")

    net = tflearn.layers.estimator.regression(net, optimizer='adam', learning_rate=0.0001, loss='categorical_crossentropy')

    model = tflearn.DNN(net, tensorboard_verbose=3)
    
    dummyData = np.reshape(np.concatenate(kfoldData[:i] + kfoldData[i+1:], axis=0), [-1, 2000, 19, 17, 1])
    dummyLabels = np.reshape(np.concatenate(kfoldLabelsOH[:i] + kfoldLabelsOH[i+1:], axis=0), [-1, 2])
    model.fit(dummyData[:,::40], dummyLabels, batch_size=8, n_epoch=30, show_metric=True)

    illTest = []
    healthTest = []
    for index, item in enumerate(kfoldLabels[i]):
        if item == 1:
            illTest.append(kfoldData[i][index])
        if item == 0:
            healthTest.append(kfoldData[i][index])

    healthLabel = np.tile([1,0], (len(healthTest), 1))
    illLabel = np.tile([0,1], (len(illTest), 1))

    sens.append(model.evaluate(np.array(healthTest)[:,::40], healthLabel))
    spec.append(model.evaluate(np.array(illTest)[:,::40], illLabel))

    # Get roc curve data
    predicted = np.array(model.predict(np.array(kfoldData[i])[:,::40]))
    fpr, tpr, th = roc_curve(kfoldLabels[i], predicted[:,1])
    roc.append([fpr, tpr])


    

In [ ]:

    

#model.save("./outData/models/3d_conv8-16-50.tflearn")
#model.load("./outData/models/3d_conv.tflearn")


    

In [ ]:

    

print("Specificity:", spec, "\nAvg:", np.mean(spec), "\nSensitivity:", sens, "\nAvg:", np.mean(sens))


    

In [ ]:

    

for i in np.arange(k):
    plt.plot(roc[i][0], roc[i][1])
plt.title("ROC curve")
plt.xlabel("FPR")
plt.ylabel("TPR")
plt.savefig("./outData/MCG/mcgKfold3D.png")