better tissue fold detection

[choosehappy]

potentially look at mode of bightness in mask to use -- tissue folds likely show 2 modes

[CielAl]

I find this hardly work in many skin dataset wherein the epidermis itself can be closed to folded regions. In fact, there are just so many different tissue components (glands/follicles) with much darker intensity in skin - are there morphological features we can use (e.g., find regions in long strip and much darker) for such tissue data?

[choosehappy]

this is a good question -- we would have to look at many examples to see if there are indeed consistent morphological features. from what I've seen there often aren't unfortunately, as the tissue folding need not happen "in a line" but can as well happen at e.g., corners

for example, this one is "kind of" linear, but not really (from the tcga: TCGA-BH-A0HO-01Z-00-DX1.D3D66547-F5D4-40F5-B737-2FECEEB35ACB.svs)

one could likely train a stain/organ specific classifier, but this goes against the idea of generalizability of histoqc. a more clever study of the patterns will be necessary to see if something is exploitable there

[jacksonjacobs1]

Here are a couple of conference papers on the subject:

Segmentation of folds in tissue section images https://ieeexplore.ieee.org/document/4353626/citations?tabFilter=papers#citations

The proposed multistage algorithm consists of three phases. First, the section image is converted to HSI color-space and the saturation and intensity components are processed in order to enhance the discrimination of the objective pixels. Next, segmentation is performed using k- means clustering and the cluster containing fold pixels is extracted from the others. Finally, unavoidable segmentation errors caused mostly by the nuclei of similar characteristics with folds are corrected based on the size and component values of the faulty segmented objects.

Note: The k-means approach seems like a fairly weak method for segmenting folds. The main takeaway is that these authors had some success with the HSI color space

---

Detection of tissue folds in whole slide images https://ieeexplore.ieee.org/document/5334529/references#references

Variation in the staining condition among tissue slides makes it complicated to choose which color channel to use for the detection of tissue folds[12]. However, the utilization of the saturation and luminance information in the enhancement procedure enables the detection of tissue folds independent from the staining condition of tissue images. From the enhanced images shown in fig. 3 it is apparent that tissue folds can be localized and segmented directly using the luminance information of the enhanced image.

[choosehappy]

nice finds, they look simple enough to implement and try out. one challenge with histoqc is that ultimately these folks likely tune specific parameters to their scanners/scanners/tissue-types/datasets, which is not something we can easily generalize to every WSI that histoqc will ever see :-\