Identifying positive pixels after color deconvolution ignoring boundaries

Question

I am analyzing histology tissue images stained with a specific protein marker which I would like to identify the positive pixels for that marker. My problem is that thresholding on the image gives too much false positives which I'd like to exclude.

I am using color deconvolution (separate_stains from skimage.color) to get the AEC channel (corresponding to the red marker), separating it from the background (Hematoxylin blue color) and applying cv2 Otsu thresholding to identify the positive pixels using cv2.threshold(blur,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU), but it is also picking up the tissue boundaries (see white lines in the example picture, sometimes it even has random colors other than white) and sometimes even non positive cells (blue regions in the example picture). It's also missing some faint positive pixels which I'd like to capture.

Overall: (1) how do I filter the false positive tissue boundaries and blue pixels? and (2) how do I adjust the Otsu thresholding to capture the faint red positives?

Adding a revised example image -

1. top left the original image after using HistoQC to identify tissue regions and apply the mask it identified on the tissue such that all of the non-tissue regions are black. I should tru to adjust its parameters to exclude the folded tissue regions which appear more dark (towards the bottom left of this image). Suggestions for other tools to identify tissue regions are welcome.
2. top right hematoxylin after the deconvolution
3. bottom left AEC after the deconvolution
4. bottom right Otsu thresholding applied not the original RGB image trying to capture only the AEC positives pixels but showing also false positives and false negatives

Thanks

Answer 1

There are several issues that cause improper quantification. I'll go over the details of how I would recommend you tackle these slides.

I'm using DIPlib, because I'm most familiar with it (I'm an author). It has Python bindings, which I use here, and can be installed with pip install diplib. However, none of this is complicated image processing, and you should be able to do similar processing with other libraries.

Loading image

There is nothing special here, except that the image has strong JPEG compression artifacts, which can interfere with the stain unmixing. We help the process a bit by smoothing the image with a small Gaussian filter.

import diplib as dip
import numpy as np

image = dip.ImageRead('example.png')
image = dip.Gauss(image, [1]) # because of the severe JPEG compression artifacts

Stain unmixing

[Personal note: I find it unfortunate that Ruifrok and Johnston, the authors of the paper presenting the stain unmixing method, called it "deconvolution", since that term already had an established meaning in image processing, especially in combination with microscopy. I always refer to this as "stain unmixing", never "deconvolution".]

This should always be the first step in any attempt at quantifying from a bightfield image. There are three important RGB triplets that you need to determine here: the RGB value of the background (which is the brightness of the light source), and the RGB value of each of the stains. The unmixing process has two components:

First we apply the Beer-Lambert mapping. This mapping is non-linear. It converts the transmitted light (as recorded by the microscope) into absorbance values. Absorbance indicates how strongly each point on the slide absorbs light of the various wavelengths. The stains absorb light, and differ by the relative absorbance in each of the R, G and B channels of the camera.

background_intensity = [209, 208, 215]
image = dip.BeerLambertMapping(image, background_intensity)

I manually determined the background intensity, but you can automate that process quite well if you have whole slide images: in whole slide images, the edges of the image always correspond to background, so you can look there for intensities.

The second step is the actual unmixing. The mixing of absorbances is a linear process, so the unmixing is solving of a set of linear equations at each pixel. For this we need to know the absorbance values for each of the stains in each of the channels. Using standard values (as in skimage.color.hax_from_rgb) might give a good first approximation, but rarely will provide the best quantification.

Stain colors change from assay to assay (for example, hematoxylin has a different color depending on who made it, what tissue is stained, etc.), and change also depending on the camera used to image the slide (each model has different RGB filters). The best way to determine these colors is to prepare a slide for each stain, using all the same protocol but not putting on the other dyes. From these slides you can easily obtain stain colors that are valid for your assay and your slide scanner. This is however rarely if ever done in practice.

A more practical solution involves estimating colors from the slide itself. By finding a spot on the slide where you see each of the stains individually (where stains are not mixed) one can manually determine fairly good values. It is possible to automatically determine appropriate values, but is much more complex and it'll be hard finding an existing implementation. There are a few papers out there that show how to do this with non-negative matrix factorization with a sparsity constraint, which IMO is the best approach we have.

hematoxylin_color = np.array([0.2712, 0.2448, 0.1674])
hematoxylin_color = (hematoxylin_color/np.linalg.norm(hematoxylin_color)).tolist()
aec_color = np.array([0.2129, 0.2806, 0.4348])
aec_color = (aec_color/np.linalg.norm(aec_color)).tolist()
stains = dip.UnmixStains(image, [hematoxylin_color, aec_color])
stains = dip.ClipLow(stains, 0) # set negative values to 0
hematoxylin = stains.TensorElement(0)
aec = stains.TensorElement(1)

Note how the linear unmixing can lead to negative values. This is a result of incorrect color vectors, noise, JPEG artifacts, and things on the slide that absorb light that are not the two stains we defined.

Identifying tissue area

You already have a good method for this, which is applied to the original RGB image. However, don't apply the mask to the original image before doing the unmixing above, keep the mask as a separate image. I wrote the next bit of code that finds tissue area based on the hematoxylin stain. It's not very good, and it's not hard to improve it, but I didn't want to waste too much time here.

tissue = dip.MedianFilter(hematoxylin, dip.Kernel(5))
tissue = dip.Dilation(tissue, [20])
tissue = dip.Closing(tissue, [50])
area = tissue > 0.2

Identifying tissue folds

You were asking about this step too. Tissue folds typically appear as larger darker regions in the image. It is not trivial to find an automatic method to identify them, because a lot of other things can create darker regions in the image too. Manual annotation is a good start, if you collect enough manually annotated examples you could train a Deep Learning model to help you out. I did this just as a place holder, again it's not very good, and identifies some positive regions as folds. Folds are subtracted from the tissue area mask.

folds = dip.Gauss(hematoxylin - aec, [20])
area -= folds > 0.2

Identifying positive pixels

It is important to use a fixed threshold for this. Only a pathologist can tell you what the threshold should be, they are the gold-standard for what constitutes positive and negative.

Note that the slides must all have been prepared following the same protocol. In clinical settings this is relatively easy because the assays used are standardized and validated, and produce a known, limited variation in staining. In an experimental setting, where assays are less strictly controlled, you might see more variation in staining quality. You will even see variation in staining color, unfortunately. You can use automated thresholding methods to at least get some data out, but there will be biases that you cannot control. I don't think there is a way out: inconsistent stain in, inconsistent data out.

Using an image-content-based method such as Otsu causes the threshold to vary from sample to sample. For example, in samples with few positive pixels the threshold will be lower than other samples, yielding a relative overestimation of the percent positive.

positive = aec > 0.1 # pick a threshold according to pathologist's idea what is positive and what is not

pp = 100 * dip.Count(dip.And(positive, area)) / dip.Count(area)
print("Percent positive:", pp)

I get a 1.35% in this sample. Note that the % positive pixels is not necessarily related to the % positive cells, and should not be used as a substitute.

Answer 2

@cris-luengo thank you for your input on scikit-image! I am one of the core developers, and based on @assafb input, we are trying to rewrite the code on color/colorconv/separate_stains.

@Assafb: The negative log10 transformation is the Beer-Lambert mapping. What I don't understand in that code is the line rgb += 2. I don't know where that comes from or why they use it. I'm 100% sure it is wrong. I guess they're trying to avoid log10(0), but that should be done differently. I bet this is where your negative values come from, though.

Yes, apparently (I am not the original author of this code) we use rgb += 2 to avoid log10(0). I checked Fiji's Colour Deconvolution plugin, and they add 1 to their input. I tested several input numbers to help on that, and ~2 would let us closer to the desirable results.

@Assafb: Compare the implementation in skimage with what is described in the original paper. You'll see several errors in the implementation, most importantly the lack of a division by the max intensity. They should have used -np.log10(rgb/255) (assuming that 255 is the illumination intensity), rater than -np.log10(rgb).

Our input data is float; the max intensity in this case would be 1. I'd say that that's the reason we don't divide by something.

Besides that, I opened an issue on scikit-image to discuss these problems — and to specify a solution. I made some research already — I even checked DIPlib's documentation —, and implemented a different version of that specific function. However, stains are not my main area of expertise, and we would be glad if you could help evaluating that code — and maybe pointing a better solution. Thank you again for your help!

Answer 3

I ended up incorporating some of the feedback given above by Chris into the following possible unconventional solution for which I would appreciate getting feedback (to the specific questions below but also general suggestions for improvement or more effective/accurate tools or strategy):

1. Define (but not apply yet) tissue mask (HistoQC) after optimizing HistoQC script to remove as much of the tissue folds as possible without removing normal tissue area
2. Apply deconvolution on the original RGB image using hax_from_rgb
3. Using the second channel which should correspond to the red stain pixels, and subtract from it the third channel which as far as I see corresponds to the background non-red/blue pixels of the image. This step removes the high values in the second channel that which up because of tissue folds or other artifacts that weren't removed in the first step (what does the third channel correspond to? The Green element of RGB?)
4. Blur the adjusted image and threshold based on the median of the image plus 20 (Semi-arbitrary but it works. Are there better alternatives? Otsu doesn't work here at all)
5. Apply the tissue regions mask on the thresholded image yielding only positive red/red-ish pixels without the non-tissue areas
6. Count the % of positive pixels relative to the tissue mask area

I have been trying to apply, as suggested above, the tissue mask on the deconvolution red channel output and then use Otsu thresholding. But it failed since the black background generated by the applying the tissue regions mask makes the Otsu threshold detect the entire tissue as positive. So I have proceeded instead to apply the threshold on the adjusted red channel and then apply the tissue mask before counting positive pixels. I am interested in learning what am I doing wrong here.

Other than that, the LoG transformation didn't seem to work well because it produced a lot of stretched bright segments rather than just circular blobs where cells are located. I'm not sure why this is happening.

Answer 4

Use ML for this case.

1. Create manually binary mask for your pictures: each red pixel - white, background pixels - black.
2. Work in HSV or Lab color space.
3. Train simple classifier: decision tree or SVM (linear or with RBF)..
4. Let's test!

See on a good and very simple example with skin color segmentation.

And in the future you can add new examples and new cases without code refactoring: just update dataset and retrain model.