Implementaton of random smooth gray value transformations (rsgt) for training gray value independent neural networks
Random Smooth Grayvalue Transformations
Convolutional neural networks trained for a detection or segmentation task in a specific type of medical gray value images, such as CT or MR images, typically fail in other medical gray value images, even if the target structure looks similar in both types of images. Random smooth gray value transformations are a data augmentation technique aimed at forcing the network to become gray value invariant. During training, the gray value of the training images or patches are randomly changed, but using a smooth and continous transfer function so that shape and texture information is largely retained.
API documentation: http://rsgt.readthedocs.io/
To use data augmentation with random smooth gray value transformations in your own project, simply install the
pip install rsgt
- Requires Python 2.7+ or 3.5+
- Numpy is the only other dependency
The expected input is a numpy array with integer values, which is usually the case for medical gray value images, such as CT and MR scans.
from rsgt.augmentation import random_smooth_grayvalue_transform # Apply gray value transformation to a numpy array new_image = random_smooth_grayvalue_transform(image, dtype='float32')
The returned numpy array will have a floating point dataype and values in the range [0,1].
While the function supports input data with any number of dimensions, it does currently not support mini-batches. A mini-batch of 3D images can be treated as a 4D input array, but all images in the mini-batch will then be subject to the same transformation. This means that a single random look up table will be computed and applied to all images in the mini-batch. There is currently no vectorized implementation of the transformation function, so a for loop is at this point the only way to transform images in a mini-batch with different transformation functions:
for i in range(minibatch.shape): minibatch[i] = random_smooth_grayvalue_transform(minibatch[i], dtype='float32')
The left most image is the original CT slice. The other images show the same slice with random smooth gray value transformations applied. The transformation function is shown below the transformed image.
Because the augmentation function returns values in the range [0,1], it is necessary to either also apply the gray value transformation at inference time, or to
normalize input images at inference time to [0,1]. The
rsgt package comes with helper functions for CT and MR scans:
Expected values of the original image are Hounsfield units ranging from -1000 for air (and below for background outside the image area) to around 3000 for very dense structures like metal implants.
from rsgt.normalization import normalize_ct_scan normalized_image = normalize_ct_scan(image, dtype='float32')
Because values of MR scans are not standardized like those of CT scans, the normalization is based on the 5% and the 95% percentiles of the input values. Values below and above are clipped.
from rsgt.normalization import normalize_mr_scan normalized_image = normalize_mr_scan(image, dtype='float32')
This normalization can also be used in combination with the augmentation technique:
from rsgt.augmentation import random_smooth_grayvalue_transform from rsgt.normalization import normalize_mr_scan N = 4096 # number of bins normalized_integer_image = (normalize_mr_scan(image, dtype='float32') * N).round().astype(int) new_image = random_smooth_grayvalue_transform(normalized_integer_image, min_max_val=(0, N), dtype='float32')
Please cite our short paper describing random smooth gray value transformations for data augmentation when using this technique in your work:
N. Lessmann and B. van Ginneken, "Random smooth gray value transformations for cross modality learning with gray value invariant networks", arXiv:2003.06158
This package is released under the MIT license, as found in the LICENSE file, with the exception of the images in the
/examples directory, which
are released under a Creative Commons license (CC BY-NC-SA 4.0).
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.