diff --git a/examples/keras_mnist_segmentation/slow_generator.py b/examples/keras_mnist_segmentation/slow_generator.py
index 1956998..6b8d098 100644
--- a/examples/keras_mnist_segmentation/slow_generator.py
+++ b/examples/keras_mnist_segmentation/slow_generator.py
@@ -1,5 +1,7 @@
 import math
 import numpy as np
+import os
+import threading
 
 from tensorflow import keras
 from unet import unet
@@ -22,7 +24,11 @@ def __getitem__(self, idx):
     batch_x = self.x[start:end, ...]
     batch_y = self.y[start:end, ...]
 
-    print(f'going to sleep for {sleep_time}s')
+    # dummy calculation to produce cpu load
+    for dd in range(10000):
+      dummy = np.random.random(100000)**2
+
+    print(f' creating batch {idx}, pid {os.getpid()}, tid {threading.get_ident()}')
     sleep(self.sleep_time)
 
     return batch_x, batch_y
@@ -32,25 +38,27 @@ def __getitem__(self, idx):
 if __name__ == '__main__':
   shape      = (800,32,32,1)
   val_shape  = (80,32,32,1)
-  batch_size = 8
-  sleep_time = 1
+  batch_size = 80
+  sleep_time = 0.1
 
   np.random.seed(1)
 
   x = np.random.random(shape)
   y = (np.random.random(shape) > 0.5).astype(float)
 
-  x_val = np.random.random(val_shape)
-  y_val = (np.random.random(val_shape) > 0.5).astype(float)
-
   gen = SLOWSequence(x, y, batch_size, sleep_time = sleep_time)
 
-  model = unet(input_shape = x.shape[1:], nfeat = 2, batch_normalization = True)
+  model = unet(input_shape = x.shape[1:], nfeat = 32, batch_normalization = True)
 
   model.compile(optimizer = keras.optimizers.Adam(learning_rate = 1e-3),
                 loss = keras.losses.BinaryCrossentropy())
 
+  # use_multiporcessing in model.fit() only works correctly in tf 2.1
+  # in tf 2.0 it is always executed in the main process
+  # in tf 2.0, use fit_generator() (which is deprecated)
   history = model.fit(gen,
-                      epochs              = 2,
-                      validation_data     = (x_val, y_val),
-                      shuffle             = False)
+                      epochs              = 4,
+                      shuffle             = False,
+                      use_multiprocessing = True,
+                      workers             = 8,
+                      max_queue_size      = 8)
diff --git a/pymirc/metrics/tf_losses.py b/pymirc/metrics/tf_losses.py
index 6c0eb8f..31823d4 100644
--- a/pymirc/metrics/tf_losses.py
+++ b/pymirc/metrics/tf_losses.py
@@ -4,7 +4,7 @@
 else:
   from tensorflow.keras import backend as K
 
-from .tf_metrics import soft_dice_coef, soft_jaccard_index, IoU
+from .tf_metrics import soft_dice_coef, soft_jaccard_index, IoU, ssim_3d
 
 
 def weighted_binary_crossentropy(weights=[.5, 1]):
@@ -66,3 +66,25 @@ def focal_loss_fixed(y_true, y_pred):
 
     return focal_loss_fixed
 
+
+def ssim_3d_loss(x, y, **kwargs):
+  """ Compute the structural similarity loss between two batches of 3D single channel images
+
+  Parameters
+  ----------
+
+  x,y : tensorflow tensors with shape [batch_size,depth,height,width,1] 
+    containing a batch of 3D images with 1 channel
+  **kwargs : dict
+    passed to tf_ssim_3d
+
+  Returns
+  -------
+  a 1D tensorflow tensor of length batch_size containing the 1 - SSIM for
+  every image pair in the batch
+
+  See also
+  ----------
+  tf_ssim_3d
+  """
+  return 1 - ssim_3d(x, y, **kwargs)
diff --git a/pymirc/metrics/tf_metrics.py b/pymirc/metrics/tf_metrics.py
index 3b11148..c7ff4df 100644
--- a/pymirc/metrics/tf_metrics.py
+++ b/pymirc/metrics/tf_metrics.py
@@ -67,6 +67,103 @@ def IoU(y_true, y_pred):
     '''
     return jaccard_index(y_true, y_pred)
 
+def tf_gauss_kernel_3d(sigma, size):
+  """Generate 3D Gaussian kernel 
+  
+  Parameters
+  ----------
+
+  sigma : float
+    width of the gaussian
+  size : int 
+    size of the gaussian (should be odd an approx 2*int(3.5*sigma + 0.5) + 1
+
+  Returns
+  -------
+  tensorflow tensor with dimension [size,size,size,1,1] with tf.reduce_sum(k) = 1
+  """
+  size  = tf.convert_to_tensor(size, tf.int32)
+  sigma = tf.convert_to_tensor(sigma, tf.float32)
+
+  coords = tf.cast(tf.range(size), tf.float32) - tf.cast(size - 1, tf.float32) / 2.0
+
+  g = -0.5*tf.square(coords) / tf.square(sigma)
+  g = tf.nn.softmax(g)
+
+  g = tf.einsum('i,j,k->ijk', g, g, g)
+  g = tf.expand_dims(tf.expand_dims(g, -1), -1)
+
+  return g
+
+def ssim_3d(x, y, sigma = 1.5, size = 11, L = None, K1 = 0.01, K2  = 0.03, return_image = False):
+  """ Compute the structural similarity between two batches of 3D single channel images
+
+  Parameters
+  ----------
+
+  x,y : tensorflow tensors with shape [batch_size,depth,height,width,1] 
+    containing a batch of 3D images with 1 channel
+  L : float
+    dynamic range of the images. 
+    By default (None) it is set to tf.reduce_max(y) - tf.reduce_min(y)
+  K1, K2 : float
+    small constants needed to avoid division by 0 see [1]. 
+    Default 0.01, 0.03
+  sigma : float 
+    width of the gaussian filter in pixels
+    Default 1.5
+  size : int
+    size of the gaussian kernel used to calculate local means and std.devs 
+    Default 11
+
+  Returns
+  -------
+  a 1D tensorflow tensor of length batch_size containing the SSIM for
+  every image pair in the batch
+
+  Note
+  ----
+  (1) This implementation is very close to [1] and 
+      from skimage.metrics import structural_similarity
+      structural_similarity(x, y, gaussian_weights = True, full = True, data_range = L)
+  (2) The default way of how the dynamic range L is calculated (based on y)
+      is different from [1] and structural_similarity()
+
+  References
+  ----------
+  [1] Wang, Z., Bovik, A. C., Sheikh, H. R., & Simoncelli, E. P.
+  (2004). Image quality assessment: From error visibility to
+  structural similarity. IEEE Transactions on Image Processing
+  """
+  if (x.shape[-1] != 1) or (y.shape[-1] != 1):
+    raise ValueError('Last dimension of input x has to be 1')
+
+  if L is None:
+      L = tf.reduce_max(y) - tf.reduce_min(y)
+    
+  C1 = (K1*L)**2
+  C2 = (K2*L)**2
+  
+  shape  = x.shape
+  kernel = tf_gauss_kernel_3d(sigma, size)
+
+  mu_x     = tf.nn.conv3d(x, kernel, strides = [1,1,1,1,1], padding = 'VALID')
+  mu_y     = tf.nn.conv3d(y, kernel, strides = [1,1,1,1,1], padding = 'VALID')
+
+  mu_x_sq  = mu_x*mu_x
+  mu_y_sq  = mu_y*mu_y
+  mu_x_y   = mu_x*mu_y
+
+  sig_x_sq = tf.nn.conv3d(x*x, kernel, strides = [1,1,1,1,1], padding = 'VALID') - mu_x_sq
+  sig_y_sq = tf.nn.conv3d(y*y, kernel, strides = [1,1,1,1,1], padding = 'VALID') - mu_y_sq
+  sig_xy   = tf.nn.conv3d(x*y, kernel, strides = [1,1,1,1,1], padding = 'VALID') - mu_x_y
+  
+  SSIM= (2*mu_x_y + C1)*(2*sig_xy + C2) / ((mu_x_sq + mu_y_sq + C1)*(sig_x_sq + sig_y_sq + C2))
+
+  if not return_image:
+    SSIM = tf.reduce_mean(SSIM, [1,2,3,4]) 
+
+  return SSIM
 
 # aliases
 # dice = DICE = dice_coef