Volume Processing in TFlow 2.x

get_mask_boundary.py

"""
This method uses the max pooling operation to get the boundary around a binary mask
Can be used in segmentation tasks where one wants to specific a loss for the boundary pixels/voxels 
"""

import os
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3" 
import nrrd
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt

# Step 1 - Define params
filename = 'nrrd_HaN_MICCAI2015-test_offsite-0522c0555_resample_True_mask.nrrd'
idx      = 40
ksize    = (6,6,3)

# Step 2 - Get boundary of mask
data, _ = nrrd.read(filename)  # [H,W,D]
data[data > 0] = 1       
data = tf.constant(data, tf.float32)     
data = tf.expand_dims(tf.expand_dims(data, axis=0), axis=-1)
data_dilated  = tf.nn.max_pool3d(data, ksize=ksize, strides=1, padding='SAME')
data_eroded   = -tf.nn.max_pool3d(-data, ksize=ksize, strides=1, padding='SAME')
data_boundary = data_dilated - data_eroded


f,axarr = plt.subplots(2,2, figsize=(10,10))
axarr[0][0].imshow(data[0,:,:,idx,0], cmap='gray')
axarr[0][0].set_title('Original Map')

axarr[0][1].imshow(data_dilated[0,:,:,idx,0], alpha=0.5, cmap='gray')
axarr[0][1].set_title('Dilated ')

axarr[1][0].imshow(data_eroded[0,:,:,idx,0], alpha=0.5, cmap='gray')
axarr[1][0].set_title('Eroded ')

axarr[1][1].imshow(data_boundary[0,:,:,idx,0], alpha=0.5, cmap='gray')
axarr[1][1].set_title('Boundary')
plt.show()

morphological_dilation.py

# Note: Chosing odd-numbered shape for kernels is important to achieve the desired dilation effect.

import numpy as np
import matplotlib.pyplot as plt
plt.imshow(np.random.random((100,100))); plt.show()

import tensorfow as tf

A = [[0, 0, 0, 0, 0, 0, 0],
     [0, 0, 0, 0, 0, 0, 0],
     [0, 0, 0, 0, 1, 0, 0],
     [0, 0, 0, 1, 1, 1, 0],
     [0, 0, 0, 0, 1, 0, 0],
     [0, 0, 0, 0, 0, 0, 0],
     [0, 0, 0, 0, 0, 0, 0]]
A  = np.expand_dims(np.expand_dims(A, axis=0),axis=-1)
A_ = tf.nn.dilation2d(A,filters=tf.zeros((3,3,1), dtype=tf.int32), strides=(1,1,1,1), padding='SAME', data_format='NHWC', dilations=(1,1,1,1))

A[0,:,:,0]
A_[0,:,:,0]

A_ = tf.nn.max_pool2D(A, ksize=(3,3), strides=1, padding='SAME')
A[0,:,:,0]
A_[0,:,:,0]