Many factors contribute to color inconsistencies in histology images, but they are primarily due to the tissue preparation and histology staining process. Other factors may include the conditions and small differences in the labs where the slides are prepared. The techniques used in the process and fixation delays as well as the conditions during slide digitization using a scanner, such as changes in light sources, detectors, or optics, contribute to the discrepancies [4]. These discrepancies in colors in the images could negatively impact the training process in CNNS. [23]. There have been many stain normalization techniques proposed. In this research, two techniques were applied, proposed by Reinhard et al. [22] and Macenko et al. [19].

These techniques aid in improving the efficiency and accuracy of a network by reducing the color inconsistencies in the images. Moreover, without stain normalization, the network may learn staining patterns instead of extracting the relevant features [14].

However, the majority of the top performing methods reported in the “ICIAR2018 Grand Challenge” paper [14] did not use any form of stain normalization, so we also conducted experiments with images that were not normalized.

Images must be converted from the BGR color space to the RGB color space in order for the stain normalization techniques to function as expected.

(1) Macenko Stain Normalization. This technique [19] accounts for the staining protocol used during the preparation of the tissue slide. Firstly, the colors are converted to optical density (OD) via the simple logarithmic transformation.

A value, β, is specified and used as a threshold to remove data with higher OD intensity. Singular value decomposition (SVD) is applied to the optical density tuples from the first step in order to determine a plane. This plane corresponds to the two largest singular values found. The optical density-transformed pixels are then projected onto this plane so that the angle at every point concerning the first SVD direction can be determined. Then, the color space transform resulting from the previous steps is applied to the original breast cancer histology image, and the histogram of the image is stretched such that the range covers the lower (100–α)% of the data. Minimum and maximum vectors are calculated and projected back into the optical density space. The hematoxylin stain corresponds to the former vector, and the eosin stain corresponds to the latter vector. The concentrations of the stains are appropriately determined, and the resulting matrix represents the RGB channels and OD intensity. The values α and β are recommended to be set to 1 and 0.15, respectively, and are kept the same for these experiments.

(2) Reinhard Stain Normalization. This technique [22] focuses on mapping the color distribution of an over- or under-stained image to a well-stained image. The use of linear transformation from RGB to lαβ color space by matching mean and standard deviation values of the color channels achieves this. Essentially, the mean color within the selected target image is transferred onto the source image. This method preserves the intensity variation of the original image. This, in turn, preserves its structure, while its contrast is adjusted to that of the target. In the lαβ color space, the stains are not precisely separated. The lαβ color space must be converted back into RGB to attain the normalized image.

Figure 4 shows examples of the stain normalization techniques applied in this study. In this figure, (a) represents the target image that was used for both techniques. Essentially, the techniques aim to normalize the colors in the original images to those of the target. An example of the original image is shown in (b). The subfigures (c) and (d) show the result of using (a) on (b) with the Macenko and Reinhard techniques, respectively.

Examples of original and normalized images: (a) target image; (b) original image; (c) Macenko-normalized; (d) Reinhard-normalized.

Note: The content above has been extracted from a research article, so it may not display correctly.

Please log in to submit your questions online.
Your question will be posted on the Bio-101 website. We will send your questions to the authors of this protocol and Bio-protocol community members who are experienced with this method. you will be informed using the email address associated with your Bio-protocol account.

We use cookies on this site to enhance your user experience. By using our website, you are agreeing to allow the storage of cookies on your computer.