stain_augment

absorbance_to_image(absorbance, source_intensity=255, dtype=torch.uint8)

Convert an absorbance (optical density) image back to a standard image. Parameters

---

ndarray

The absorbance image to convert back to a linear intensity range.

float, optional

Reference intensity for image. This should match what was used with rgb_to_absorbance when creating absorbance.

numpy.dtype, optional

The datatype to cast the output image to.

Returns

ndarray

An image computed from the absorbance

image_to_absorbance(image, source_intensity=255.0, dtype=torch.float32)

Convert an image to units of absorbance (optical density). Parameters

---

ndarray

The image to convert to absorbance. Can be single or multichannel.

float, optional

Reference intensity for image.

numpy.dtype, optional

The floating point precision at which to compute the absorbance.

Returns

ndarray

The absorbance computed from image.

Notes

If image has an integer dtype it will be clipped to range [1, source_intensity], while float image inputs are clipped to range ``[source_intensity/255, source_intensity]. The minimum is to avoid log(0). Absorbance is then given by .. math:: absorbance = \log{\frac{image}{source_intensity}}.

stain_extraction_pca(image, source_intensity=240, alpha=1, beta=0.345, ref_stain_coeff=((0.5626, 0.2159), (0.7201, 0.8012), (0.4062, 0.5581)), channel_axis=0, image_type='intensity')

Extract the matrix of H & E stain coefficient from an image. Uses a method that selects stain vectors based on the angle distribution within a best-fit plane determined by principle component analysis (PCA) [1]_. Parameters

---

cp.ndarray

RGB image to perform stain extraction on. Intensities should typically be within unsigned 8-bit integer intensity range ([0, 255]) when image_type == "intensity".

float, optional

Transmitted light intensity. The algorithm will clip image intensities above the specified source_intensity and then normalize by source_intensity so that image intensities are <= 1.0. Only used when image_type=="intensity".

float, optional

Algorithm parameter controlling the [alpha, 100 - alpha] percentile range used as a robust [min, max] estimate.

float, optional

Absorbance (optical density) threshold below which to consider pixels as transparent. Transparent pixels are excluded from the estimation.

Additional Parameters

int, optional

The axis corresponding to color channels (default is the last axis).

{"intensity", "absorbance"}, optional

With the default image_type of "intensity", the image will be transformed to absorbance units via image_to_absorbance. If the input image is already an absorbance image, then image_type should be set to "absorbance" instead.

Returns

cp.ndarray

Stain attenuation coefficient matrix derived from the image, where the first column corresponds to H, the second column is E and the rows are RGB values.

Notes

The default beta of 0.345 is equivalent to the use of 0.15 in [1]_. The difference is due to our use of the natural log instead of a decadic log (log10) when computing the absorbance. References

---

.. [1] M. Macenko et al., "A method for normalizing histology slides for quantitative analysis," 2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2009, pp. 1107-1110, doi: 10.1109/ISBI.2009.5193250. http://wwwx.cs.unc.edu/~mn/sites/default/files/macenko2009.pdf