metrics.py

from keras import backend as K
import tensorflow as tf
import numpy as np


def dice_coefficient(y_true, y_pred, smooth=0.1):
    y_true_f = K.flatten(y_true)
    y_pred_f = K.flatten(y_pred)

    intersection = K.sum(y_pred_f * y_true_f)
    return (2. * intersection + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)


def dice_coefficient_loss(y_true, y_pred):
    return -dice_coefficient(y_true, y_pred)

def sigmoid(x):
    return 1. / (1. + K.exp(-x))

def weighted_crossentropy_pixelwise(y_true, y_pred):

    y_pred = tf.clip_by_value(y_pred, 1e-7, 1 - 1e-7)
    y_pred = K.log(y_pred / (1 - y_pred))

    wmh_indexes = np.where(y_true == 1.0)
    weights = np.repeat(1.0, 240 * 240)
    weights = np.reshape(weights, (1, 240, 240, 1))
    weights[wmh_indexes] = 5000.0

    crossentropy = (y_true * weights * -K.log(sigmoid(y_pred)) + (1 - y_true * weights) * -K.log(1 - sigmoid(y_pred)))
    return crossentropy


def prediction_count(y_true, y_pred):
    return tf.count_nonzero(y_pred)


def label_count(y_true, y_pred):
    return tf.count_nonzero(y_true)


def prediction_sum(y_true, y_pred):
    return tf.reduce_sum(y_pred)


def label_sum(y_true, y_pred):
    return tf.reduce_sum(y_true)


dice_coef = dice_coefficient
dice_coef_loss = dice_coefficient_loss

weighted_crossentropy = weighted_crossentropy_pixelwise

predicted_count = prediction_count
predicted_sum = prediction_sum

ground_truth_count = label_count
ground_truth_sum =   label_sum