Augmentation-PyTorch-Transforms

Image data augmentation on-the-fly by adding new class on transforms in PyTorch and torchvision.

Normally, we from torchvision import transforms for transformation, but some specific transformations (especially for histology image augmentation) are missing.

Thus, we add 4 new transforms class on the basic of torchvision.transforms pyfile, which we named as myTransforms.py.

You can call and use it in the same form as torchvision.transforms. Or, you can refer to dataAug_myTransforms.py.

Also, you can check the actual effect of myTransforms for data augmentation :)

New transforms classes included in myTransforms

HEDJitter

Randomly perturbe the HED color space value on an RGB pathological image[1].

1. Disentangle the hematoxylin and eosin color channels by color deconvolution[2] method using a fixed matrix.
2. Perturbe the hematoxylin, eosin and DAB stains independently.
3. Transform the resulting stains into regular RGB color space.

Args

• theta (float): How much to jitter HED color space,
• then, alpha is chosen from a uniform distribution [1-theta, 1+theta]
• betti is chosen from a uniform distribution [-theta, theta]
• the jitter formula is $s' = \alpha * s + \betti$

Example

import myTransforms
imagename = '../data/10-05074_353_49_8178.png'
img = Image.open(imagename) # read the image
	
preprocess = myTransforms.HEDJitter(theta=0.05)
print(preprocess)
	
HEPerimg = preprocess(img)
plt.subplot(121)
plt.imshow(img)
plt.subplot(122)
plt.imshow(HEPerimg)
plt.show()

References
[1]. Tellez, D., Balkenhol, M., Otte-Höller, I., van de Loo, R., Vogels, R., Bult, P., ... & Litjens, G. (2018). Whole-slide mitosis detection in H&E breast histology using PHH3 as a reference to train distilled stain-invariant convolutional networks. IEEE transactions on medical imaging, 37(9), 2126-2136.
[2]. Ruifrok, A. C., & Johnston, D. A. (2001). Quantification of histochemical staining by color deconvolution. Analytical and quantitative cytology and histology, 23(4), 291-299.

RandomElastic

Random Elastic transformation by CV2 method on image by alpha, sigma parameter.
WARNING: This transform class will spend a lot of CPU time for preprocessing.

Args

• alpha (float): alpha value for Elastic transformation, factor on dx, dy
  -- if alpha is 0, output is the same as origin, whatever the sigma;
  -- if alpha is 1, output only depends on sigma parameter;
  -- if alpha < 1 or > 1, it zoom in or out the sigma-relevant dx, dy.
• sigma (float): sigma value for elastic transformation, should be $\in (0.05,0.1)$
• mask (PIL Image) For processing on GroundTruth of segmentation task, if not assign, set None. only used in __call__ function

Example

import myTransforms
imagename = '../data/10-05074_353_49_8178.png'
img = Image.open(imagename) # read the image
	
preprocess = myTransforms.RandomElastic(alpha=2, sigma=0.06)
print(preprocess)
	
elasticimg = preprocess(img, mask=None)
plt.subplot(121)
plt.imshow(img)
plt.subplot(122)
plt.imshow(elasticimg)
plt.show()

References
affine and elastic transform
cv2.warpAffine
scipy.ndimage.map_coordinates

RandomAffineCV2

Random Affine transformation by CV2 method on image by alpha parameter.
It is different from torchvision.transforms.RandomAffine, which is implemented by PIL.Image method. We can set BORDER_REFLECT for the area outside the transform in the output image while original RandomAffine can only fill by a specified value.

Args

• alpha (float): alpha value for affine transformation
• mask (PIL Image) For processing on GroundTruth of segmentation task, if not assign, set None.

Example

import myTransforms
imagename = '../data/10-05074_353_49_8178.png'
img = Image.open(imagename) # read the image
	
preprocess = myTransforms.RandomAffineCV2(alpha=0.1)#alpha \in [0,0.15]
print(preprocess)
	
affinecvimg = preprocess(img)
plt.subplot(121)
plt.imshow(img)
plt.subplot(122)
plt.imshow(affinecvimg)
plt.show()

RandomGaussBlur

Random Gauss Blurring on image by radius parameter. Args

• radius (list, tuple): radius range for selecting from; you'd better set it < 2 especially for histopathological image task.

Example

import myTransforms
imagename = '../data/10-05074_353_49_8178.png'
img = Image.open(imagename) # read the image
	
preprocess = myTransforms.RandomGaussBlur(radius=[0.5, 1.5])
print(preprocess)
	
blurimg = preprocess(img)
plt.subplot(121)
plt.imshow(img)
plt.subplot(122)
plt.imshow(blurimg)
plt.show()

AutoRandomRotation

change torchvision.transforms.RandomRotation for auto-random select angle from [0, 90, 180, 270] for rotating the image.

Example

import myTransforms
imagename = '../data/10-05074_353_49_8178.png'
img = Image.open(imagename) # read the image
	
preprocess = myTransforms.AutoRandomRotation()
print(preprocess)
	
rotateimg = preprocess(img)
plt.subplot(121)
plt.imshow(img)
plt.subplot(122)
plt.imshow(rotateimg)
plt.show()