Hematoxylin and Eosin Stained Histopathological Image

Enhancement Method

Bogusław Cyganek

AGH University of Krakow, Al. Mickiewicza 30, 30-059 Kraków, Poland

Keywords: Medical Image Processing, Histopathology, H&E Staining, Whole Slide Images (WSI),

High-Dynamic Range, Image Enhancement, Deep Learning.

Abstract: Hematoxylin and eosin staining is one of the most well-known and common methods of staining

histopathological samples. Its main purpose is to highlight the morphological features of tissues, which help

doctors make the right diagnosis. However, it is not without its flaws, and the scans obtained in this way are

characterized by high inconsistency not only resulting from the variability of the tissues themselves, but also

due to the chemical reagents used, the technique of preparing the preparation, etc. This causes various

difficulties and errors in the case of tissue assessment performed by the algorithm, but can also be a hindrance

for doctors. Therefore, there are many methods to improve the quality of scans obtained from tissue stained

in the H&E way. In this article, we present a fairly recent idea and very preliminary results for the use of our

multi-channel virtual high-dynamic range MVHDR method to improve the parameters of H&E scans. Our

method allows both data augmentation for CNN, but also significant detail enhancement that helps doctors

identify the disease.

1 INTRODUCTION

The examination of a tissue sample is one of the basic

medical procedures in the diagnosis of various types

of cancer in humans and animals. For this purpose,

there are strict procedures for collecting such tissues

and then properly preparing, storing and - recently -

digitizing them into so called Whole Slide Images

(WSI) (Tellez 2019, Janowczyk 2019, Greeley 2024).

One of the very important steps in this chain is the

appropriate staining of the tissue. Hematoxylin and

eosin stain (H&E) is one of the principal tissue stains

used in histology for over a century (Tellez 2019).

Hematoxylin is a natural chemical compound

obtained from the logwood tree Hematoxylon

campechianum, discovered by the Spanish during

exploration expeditions to Yucatan, Mexico, in the

early 16th century. Hematoxylin's initial uses were

for dyeing hair and fabrics. The first tissue staining

with hematoxylin dates back to 1800. Interestingly,

for over 200 years, hematoxylin has remained one of

the primary methods of staining tissue, primarily for

the isolation of nuclei in microscope slides (Titford

2005). Despite many years and the development of

other methods, H&E is the most widely used stain in

biology and medical diagnosis and is often the gold

standard (Dapson 2009)(Bassotti 2011)(Ma 2024).

For example, when a pathologist looks at a biopsy of

a suspected cancer, the histological section is likely to

be H&E stained (Sorenson 2014)(Srinidhi 2021).

Figure 1: Examples of the healthy prostate tissue (upper

row) and tissues with cancer of type Gleason 5 (lower row).

Scans from the DiagSet dataset (Koziarski 2024).

Figure 1 shows patches extracted from a WSI scan

of the DiagSet dataset containing examples of

prostate tissues from the anonymous patients

(Koziarski 2024). Upper row contains the healthy

378

Cyganek, B.

Hematoxylin and Eosin Stained Histopathological Image Enhancement Method.

DOI: 10.5220/0013786000003982

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 22nd Inter national Conference on Informatics in Control, Automation and Robotics (ICINCO 2025) - Volume 2, pages 378-385

ISBN: 978-989-758-770-2; ISSN: 2184-2809

prostate tissues, while the lower one shows patches

with a prostate cancer classified by an expert to

Gleason grade no. 5.

As already mentioned, H&E uses two histological

stains – hematoxylin (H) and eosin (E). H affects cell

nuclei and renders them purplish blue. On the other

hand, E stains in pink mostly the extracellular matrix

and cytoplasm, whereas other structures take on

different shades, hues, as well as combinations of

similar colors. A brief explanation of the main

structures visible in a tissue sample after the H&E

staining is shown in Figure 2 (Lisowski 2019)

(Sampias 2025). Hence a pathologist can easily

differentiate between the nuclear and cytoplasmic

parts of a cell, and additionally, the overall patterns of

coloration from the stain show the general layout and

distribution of cells and provides a general overview

of a tissue sample's structure. Thus, pattern

recognition, both by expert humans themselves and

by software that aids those experts in digital

pathology, provides histologic information.

Figure 2: Brief explanation of the main structures visible in

a tissue sample after the hematoxylin & eosin staining

(Source images: Mikael Häggström. Public Domain, 2022).

Nevertheless, H&E is not a remedy for all the

cases and tissues – hence, in some difficult cases

more specific stains and methods are used. Similarly,

when using H&E stained patches for grading we may

encounter a number of problems. The first problem is

the heterogeneity of samples even those performed in

the same laboratory. The second is the insufficient

quality parameters of the image that make it difficult

or negatively affect the quality of classification.

In this paper we address at least two of these

problems by proposing a new concept of nonlinear

multi-channel virtual high dynamic range (MVHDR)

filtering method, however successfully applied to the

H&R scans. What's important about our method is

that it follows the general idea of tissue staining –

because unstained tissues lack contrast, our method

goes a step further and significantly improves contrast

in already H&E stained tissues.

This our preliminary proof of concept is based on

a number of previous works and discoveries that

inspired us to use it to improve the quality of H&E

scans based on the advance image filtering algorithms

(Grabek 2019)(Koziarski 2018). These previous

observations include the application of VHDR to

thermal images, i.e. from the far infrared, for which

we observed that the application of VHDR before

using the CNN network significantly improved its

results (Knapik 2019). Early concepts and results,

presented in this paper, show that our assumptions are

correct, although we leave more in-depth studies for

the future.

This is a position paper showing our initial

concepts and the first observations. The rest of the

paper is organized as follows. In Section 2, we briefly

review the existing literature on this topic. In Section

3, we present the details of our approach, with a novel

multi-channel VHDR (MVHDR). Chapter 4 presents

the results of the experiments along with a discussion

of the obtained results. The article ends with

conclusions and a list of literature.

2 RELATED WORKS

The history of the discovery of hematoxylin, as well

as its subsequent uses, is particularly interesting

(Titford 2005)(Cooksey 2021). Despite numerous

studies and new approaches to tissue contrast, H&E

remains one of the most widely used methods.

However, it is not without its drawbacks (Dapson

2009)(Bassotti 2011)(Ma 2024). Hence, there are

many works and studies on improving the image

quality of scans stained in the H&E method. Here we

mention only a few of the most important ones, and

further literature references can be found in the

mentioned here works. The first group are methods

improving the repeatability and homogenity of

staining, because even the same laboratories can

produce completely different shades for the same

tissues. These are called stain normalization methods.

In this respect Janowczyk et al. proposed the use of

sparse autoencoders for stain normalization

(Janowczyk 2017). In their method pixels are

separated into 𝑘 clusters. Then histogram equalization

across clusters and RGB channels is applied to obtain

a color standardized image. In the same vein Zanjani

et al. propose to use the deep generative models

(Zanjani 2018). This is done to separate pixels into 𝑘

tissue classes. In the next step, stain normalization is

obtained by separation of the source and target

Hematoxylin and Eosin Stained Histopathological Image Enhancement Method

379

images. On the other hand, for stain normalization

Tellez et al. propose an U-Net-like network

architecture. This is improved with heavily color-

augmented images and trained to reconstruct their

original appearance (Tellez 2019). Their main idea is

that when trained with images from a target center,

the network should be able to transform new images

to the same target color distribution.

The second group of methods deals with the

general change of parameters of H&E images. An

interesting approach is based on Blind Color

Deconvolution (BCD) techniques. Its idea is to

separate H&E images into colors (stains) and

structural information (concentrations). This, in turn,

can be useful for the further processing, data

augmentation, and classiﬁcation etc. In this respect,

Ruifrok et al. proposed the use of the logarithmically

inverted optical density space and a non-blind color

deconvolution algorithm to obtain the stain

concentrations (Ruifrok 2001).

However, we have adopted a slightly different

approach, which involves both improving contrast

and the ability to change the color space.

3 METHOD DESCRIPTION

In this section the basic architecture of the proposed

method is presented. More concretely, we start with

the overall view of the main blocs, after which a

general description of the method operation follows.

3.1 Operation of the Virtual

High-Dynamic Range Converter

The main idea of the presented method, called VHDR

(Knapik 2021), is to increase the dynamics of the

image, but using a single image as its input – hence

the name "virtual" high dynamic range. This is

different from classic HDR, which usually uses

several images with different exposures (Sen 2016).

However, in many cases, such as H&E, we simply do

not have such many exposures. Hence the idea that

we can "artificially" generate them. The processing

chain leading to this is shown in Figure 3. It operates

as follows. An input image is processed by a set of

tone adjustment curves. As a result, a number of tone

converted images is calculated.

The main idea here is to expose different ranges

of the input image, in order to reveal not well visible

details. The tone sub-images are then joined back to

form one HDR image. After that, image range

conversion and contrast enhancement are applied.

Tone curves Fusion module

Contrast

enhancement filter

Image range

conversion

image

Out

image

Figure 3: Architecture of the virtual high-dynamic range

converter. Only a single image is required as its input.

The luminance change is done with help of the

logistic function, in its basic form given as follows

()

−−

kx x

sx A e

(2)

where A denotes and amplitude and k is a parameter;

denotes the so called middle point of the S shape

function, whose inflection point is just x

Figure 4: Sets of functions to generate virtual exposures.

Horizontal axis denotes the input brightness. Vertical the

normalized output brightness. From top: convex and

concave curves.

However, to generate series of virtual exposures

we need three variants of the logistic functions:

convex, concave, as well as S-shaped. Exemplary

variants of these function are shown in Figure. Not

less important is to assure that the domain and

codomain cover the whole allowable range, which for

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

380

8-bits runs 256 values. Hence (1) needs to be

conditioned, as follows

()

,, 1

−

sxkx A e

(2)

D is a domain range constant, A=P

max

is a maximal

pixel range of the input image. The parameter D

controls the range of values of s

– it needs to be

different from 0 and 1. In other words, these are the

saturation values. In our system this parameter is set

to 2 times 6 around x

, and hence D=12.

The above function is extended by the additional

multiplicative and additive components to provide a

concave, convex and S like shape. It needs also to

cover full range of the codomain – this is in the range

min

, P

max

], i.e. [0,255]. Hence, new version of ()

reads as follows:

()

100 10

,, 1 0,,

−++

mid mid

sxkx

xkx x P P s kx

(3)

For 8-bit pixel representations P

mid

=(P

max

min

)/2=(255.0-0.0)/2=127.5, which denotes a middle

value of the low dynamic image. In the next step,

parameter x

is chosen from the following three

values

{

}

min mid max

xPPP∈

(4)

Each of the above values of x

corresponds to a

distinct type of a curve as follows (Figure 4):

1. Convex case – The midtone data of the input

image is stretched toward the highlight of the

output space, resulting in a bright-toned image.

2. Concave case – The midtone data of the input is

stretched toward the shadow of the output space,

resulting in a dark-toned image.

3. S-curve – The highlight and shadow of the input

pixels are enhanced, leading to a higher contrast.

The parameter to consider now is k, whose value

was chosen experimentally, as follows:

{

}

{}

0.2, 0.3 ,

0.4, 0.5 ,

0.3, 0.4 , .

=



∈=







min

mid

max

or x P

kforxP

or x P

(5)

Summarizing, with three values x

, and two

parameters k for each x

, a series of V=6 tone

converted images I

is generated. Values of these

parameters were chosen experimentally.

On the other hand, the fusion module takes as its

input the aforementioned set of V tone mapped

images. In its next step, each pixel i

in the output

fused image I

is computed as a weighted sum of all

V pixels i

, as follows

11==



hvvv

iwiw

(6)

where i

denotes an output pixel, i

is an input pixel

from one of the V tone mapped images I

, while the

weight w

is computed as follows:

vmid

mid



−





(7)

λ in the above denotes a parameter that controls

steepness of the weighting function. In our

experiments λ is set to a value in the range 2.2-3.9.

Pixel values obtained thanks to (6) are frequenlty

concentrated around the middle values of the

allowable pixel range, since these are increased by the

weights in (7). Therefore, their values need to be

further scaled by the image range conversion module,

in order to spread equally their histograms. This

scaling process is done as follows:

()

() ()

max

iminI

max I min I

−

, for each

iI∈

(8)

where

min(I

) and max(I

) are minimal and maximal

values of the whole I

image, respectively.

The last step constitutes image contrast

enhancement. This is done with the one-scale local

spatial filter (Cvetkovic, 2007). In this process, each

pixel is converted as follows:

()

ˆˆ ˆ





=+ −





oh h h

ii imi

(9)

where the mean value

m around a pixel

, in the

window controlled by two parameters

and t

, is

computed in accordance with the following formula

()

()()

()

2121

=− =−



sstt

mi pq

ipsqt

(10)

Figure 4 shows different functions for generation of

virtual exposures.

3.2 Multi-Channel VHDR

The VHDR method described in the previous chapter

has been extended to the space of color images or

even multi-channel signals/images. This is one of the

main contributions of this paper.

We assume that the input color image is

represented using the RGB space. Then, an optional

conversion of this space to another color space takes

place. Out tested transformations are as follows:

Hematoxylin and Eosin Stained Histopathological Image Enhancement Method

381

1. RGB  HSI

2. RGB  IJK

3. RGB  YCrCb

In each of the above spaces, including RGB, the

image dynamics improvement is performed

independently in each of the color channels. This

means that it is possible to independently select the

parameters of the single-channel VHDR method for

each color channel independently. In this way, for

example, after transformation to the YCrCb space,

only the Y channel can be modified, leaving Cr and

Cb unchanged, etc. Operation of the proposed

extended method is depicted in Figure 5.

Jednokanałowy

VHDR

Color image

in RGB

Color-VHDR

Color Space

Converter from

RGB

Jednokanałowy

VHDR

Single-channel

VHDR

Color Space

Converter to RGB

Figure 5: Architecture of the proposed multi-channel

VHDR method. Structure of the single-channel VHDR is

shown in Figure 3.

It is also possible to change all channels, also

causing a change in the color palette. The last feature

can be independently useful for doctors for deeper

analysis of histopathological scans. Selected

experimental results are presented in the next section.

4 EXPERIMENTAL RESULTS

As alluded to previously, in this position paper we

present our novel concept and intuition, as well as

initial findings. Our method is also underpinned with

good results obtained in thermal image enhancement

when operating with CNN (Knapik 2021). The main

experiment planned for the nearest future research

into this topic will be to train CNN for cancer

classification with MVHDR on its input. This

requires finding the optimal hyperparameters of

MVHDR. Hence, the next step can be to built-in

MVHDR as an initial layer of a CNN, which will be

then trained with the optimization criteria of the best

accuracy. However, this is left for the future research.

An interesting and parallel branch of method

assessment is its application in computer systems

aimed at helping to diagnose or to annotate WSI by

the professional pathologists. Based on our private

conversations with the pathologists we know their

needs. We can also assess the usefulness of such

system features as the increase in scan quality

presented here or the change of the color palette for

the histopathological diagnostic process. It will

therefore be a human-in-the-loop system.

Following this direction, in this paper we present

results comparing the quality of patches from WSI

scans before and after applying MVHDR. Our

objective measures are sharpness and clarity.

• Sharpness – a measure conveying information

on level of detail in a scan. In our approach we

compute variance of the Laplacian, defined as

follows (Cyganek 2009):

() ()

∂∂



=∇ = +



∂∂

LIxy Ixy

(11)

In our experiments a discrete version of

L is

used, to compute sharpness coefficient S over an

image I, as follows:

[

]

()

var











SLII

(12)

where

()



L I

denotes discrete Laplacian over

the entire image I, and

var stands for variance.

Hence, this measure indicates how well the

edges and fine details are captured. In many

classification systems, also based on AI,

sharpness can be essential to obtain high

accuracy especially when processing images

with some lighting and/or geometrical defects.

• Clarity – a measure that encompasses the overall

visual coherence and an overall level of noise

and/or distortions. In our approach clarity

C is

expressed as a product of the above sharpness

measure

S, and the standard deviation of an

image. Hence, clarity is expressed as follows:

[

]

[

]

[

]

=⋅CSstdII I

(13)

where

std is the standard deviation.

Figure 6 shows a number of original H&E patches

from the DiagSet with various Gleason degrees

(Koziarski 2024), i.e. these are cancerous tissues.

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

382

Figure 6: Left column – examples of prostate scan sections

at 40x magnification from DiagSet – diseased tissue – G5

on the Gleason scale. Right column – images processed

with the developed MVHDR module. Visible improvement

in contrast enabling analysis of morphological structures of

the tissue, such as estimation of the size and locations of

nucleoli.

In Figure 6 original patches are in the left column,

whereas MVHDR processed in the right one. Even

visual inspection shows that the latter are sharper and

more revealing of the morphological details of the

tissue, which may be important both for

classification, as well as for inspection by an expert

pathologist. These observations are confirmed by

quantitative measurements of the parameters S from

(12) and C defined in (13), and presented in Table 1

and Table 2, respectively.

Table 1: Sharpness values S computed for the images from

Figure 6.

Image no. S - Original S - MVHDR Ratio

1 148.27 2465.83 16.6

2 155.66 2398.71 15,4

3 108.11 1690.45 15,6

4 219.39 3246.61 14,8

5 288.54 3376.86 11,7

Figure 7 shows prostate cancer tissues showing

color space change obtained on histopathological

scans using the developed MVHDR method. Original

H&E patches are shown in the left column, while

MVHDR-processed image are in the right one.

Figure 7: Samples of the prostate cancer tissues illustrating

image enhancement and color space change by the

MVHDR method. Original H&E image (left column),

MVHDR-processed image (right column).

As can be seen in the right column of Figure 7, the

histological images have not only been contrast-

enhanced, but also converted to a different color space,

which allows for the perception of other image details,

e.g., in the case of a diagnosis made by a pathologist.

Table 2: Clarity values C computed for the images from

Figure 6.

Image no. C- Original C - MVHDR Ratio

1 6957.79 184386.41 26.5

2 4143.99 128363.72 30.9

3 4065.20 104402.24 25.7

4 8071.64 212568.42 26.3

5 16212.25 281219.42 17.3

Hematoxylin and Eosin Stained Histopathological Image Enhancement Method

383

Table 3 and Table 4 show numerical results of the

parameters S and C computed for the images from

Figure 7, respectively.

Table 3: Sharpness values S computed for the images from

Figure 7.

Image no. S - Original S - MVHDR Ratio

1 597.25 4647.94 7.8

2 301.59 3412.69 11.3

3 288.05 3777.61 13.1

Table 4: Clarity values C computed for the images from

Figure 7.

Image no. C- Original C - MVHDR Ratio

1 12583.80 188048.80 14.9

2 8096.60 188842.25 23.3

3 8106.07 189127.10 23.3

In all cases we see a significant difference, i.e.

more than an order of magnitude, between the

parameters for the original H&E scans and their

versions processed with our MVHDR method.

However, it's not the numerical values themselves

that matter, but their ratio (the rightmost column),

which in all cases exceeds an order of magnitude.

This demonstrates significant potential for improving

contrast, as well as the dynamics of pixel

representation itself. We have presented results here

for several images, but they are consistent for all

patches obtained from the WSI scan. This proves the

stability of the proposed method, which in virtually

every real case leads to a significant contrast

enhancement and – as we have seen – also allows for

changing the color palette.

The presented method was implemented in C++.

Experiments were conducted on a computer with 128

GB of RAM and an Intel® i7-11850H/2.50GHz

microprocessor, running Windows 10 Pro. Such

implementation allows for real-time processing of

video streams, as well as for easy parallelization – the

feature which we intend to utilize in the future.

5 CONCLUSIONS

In this paper we address the problem of enhancement

and filtering of the H&E stained histopathological

scans. Our original concept relies on application of

the nonlinear multi-channel virtual high dynamic

range filtering method to the H&R scans.

We would like to stress that this is an initial

concept describing work in progress and published as

a position paper. Therefore we only outlined the main

concept and presented the initial results, which are

encouraging. Also, our previous experience with this

type of data preprocessing/augmentation for CNN

training, although tested for thermal images showed

very good results, increasing the final accuracy by

certain percentage points (Knapik 2021).

Summarizing, our proposed MVHDR method

can be used for:

• Data augmentation for CNN/ViT training.

• Generation of new images for GAN.

• To generate better quality histopathological

images for doctors’ diagnosis.

Our scientific hypothesis and things to do are as

follows:

1. The proposed method, when used as a data

preprocessing module, can lead to higher

accuracies in deep learning with CNN and ViT.

2. The proposed method can be used as a data

augmentation module.

3. The method can be used to improve quality

and/or change the color palette when used in

tissue diagnosis by medical experts.

If the above are correct, then the next step would be

as follows:

4. Design of the input layers, so the optimal

parameters of the MVHDR method can be

learned by a CNN/ViT during its training.

Finally, we would like to mention that this method

was discussed and pre-tested by two pathology

experts who gave it a positive review. They were

particularly interested in the possibility of enhancing

the contrast and highlighting important

morphological details of the observed tissues. We are

planning further tests in this direction as well. It

would be particularly interesting to observe

differences in the labeling process of different types

of tissue by medical specialists without and with the

use of the method proposed here.

ACKNOWLEDGMENT

Research under project no. 10741, supported by

program “Excellence initiative – research university”

for the AGH University of Krakow.

ICINCO 2025 - 22nd International Conference on Informatics in Control, Automation and Robotics

384

REFERENCES

Bassotti, G., Villanacci, V., Salerni, B., Maurer, C. A.,

Cathoma, G. (2011). Beyond hematoxylin and eosin:

the importance of immunohistochemical techniques for

evaluating surgically resected constipated patients

Tech. Coloproctol 15:371–375, Springer, DOI

10.1007/s10151-011-0721-5.

Cooksey, C. J. (2021) Hematoxylin in the 21st century,

Biotechnic & Histochemistry, 96:3, pp. 242-249, DOI:

10.1080/10520295.2020.1786725

Cvetkovic S.D., Schirris J., de With P.H.N. (2007).

Locally-Adaptive Image Contrast Enhancement

without Noise and Ringing Artifacts. IEEE

International Conference on Image Processing, vol. 3,

pp. 551-560.

Cyganek, B. (2009). An Introduction to 3D Computer

Vision Techniques and Algorithms, Wiley,.

Dapson RW., Horobin RW. (2009). Dyes from a twenty-

first century perspective, Biotechnic & Histochemistry,

84:4, pp. 135-137, DOI: 10.1080/10520290902908802

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn,

D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer,

M., Heigold, G., Gelly, S. (2020). An image is worth

16x16 words: Transformers for image recognition at

scale. arXiv preprint. arXiv: 2010.11929.

https://doi.org/10.48550/arXiv.2010.11929.

Grabek, J., Cyganek, B (2019). Speckle Noise Filtering in

Side-Scan Sonar Images Based on the Tucker Tensor

Decomposition, Sensors, 19(13), 2903; https://doi.org/

10.3390/s19132903

Greeley, C., Holder, L., Nilsson, E. E., Skinner, M. K.

(2024). Scalable deep learning artificial intelligence

histopathology slide analysis and validation. Scientific

Reports, 14(1), 26748. https://doi.org/10.1038/s41598-

024-76807-x.

Janowczyk A, Madabhushi A. (2016). Deep learning for

digital pathology image analysis: a comprehensive

tutorial with selected use cases. J Pathol

Inform;7(1):29.

Janowczyk, Basavanhally, A., Madabhushi A. (2017).

Stain normalization using sparse autoencoders

(stanosa): application to digital pathology, Comput.

Med. Imaging Graph. 57.

Ke W.-M., Wang T.-H., Chiu C.-T. (2009). Hardware-

efﬁcient virtual high dynamic range image

reproduction. Proceedings of the 16th IEEE

International Conference on Image Processing

(ICIP’09). Piscataway, NJ, USA: IEEE Press, pp.

2665–2668.

Knapik M., Cyganek B. (2021). Fast eyes detection in

thermal images. Multimedia Tools and Applications

80:3601–3621, Springer, https://doi.org/10.1007/

s11042-020-09403-6.

Koziarski, M., Cyganek, B. (2018). Marine Snow Removal

Using a Fully Convolutional 3D Neural Network

Combined with an Adaptive Median Filter, 24th

International Conference on Pattern Recognition,

Pattern Recognition And Information Forensics (ICPR

2018) LNIP Vol. 11188 , pp. 16-25

Koziarski, M., Cyganek, B., Niedziela, P., Olborski, B.,

Antosz, Z., Żydak, M., Kwolek, B., Wąsowicz, P.,

Bukała, A., Swadźba, J. (2024). DiagSet: A dataset for

prostate cancer histopathological image classification.

Scientific Reports, 14(1), 6780. Nature Publishing

Group. https://doi.org/10.1038/s41598-024-52183-4.

Lisowski, A. (2019) Science of H&E. Leica BioSystems

https://www.leicabiosystems.com/knowledge-

pathway/science-of-he/

Ma, ZY, Zhang, XF, Hu, YZ, Zhu, MD, Jin, J, Qian, P.

(2024). Comparison of staining quality between rapid

and routine hematoxylin and eosin staining of frozen

breast tissue sections: an observational study. J Int Med

Res. Jun; 52(6): 3000605241259682. doi:

10.1177/03000605241259682. PMID: 38886869;

PMCID: PMC11184997.

Ruifrok A.C., Johnston D.A. (2001). Quantiﬁcation of

histochemical staining by color deconvolution, Anal.

Quant. Cytol. Histol. 23, pp. 291–299.

Sampias, C., Rolls, G. (2025). H&E Staining Overview: A

Guide to Best Practices. Leica BioSystems.

https://www.leicabiosystems.com/knowledge-

pathway/he-staining-overview-a-guide-to-best-

practices/

Sen P., Aguerrebere C. (2016). Practical High Dynamic

Range Imaging of Everyday Scenes. IEEE Signal

Processsing Magazine, pp. 36-44.

Srinidhi, C.L., Ciga, O., Martel, A.L. (2021). Deep neural

network models for computational histopathology: A

survey, Medical Image Analysis, Volume 67, 2021,

101813, ISSN 1361-8415, https://doi.org/10.1016/j.

media.2020.101813.

Sorenson, R.L., Brelje, T.C. (2014). Atlas of Human

Histology A Guide to Microscopic Structure of Cells,

Tissues and Organs 3rd Edition

Tellez D, Litjens G, Bándi P. (2019). Quantifying the

Effects of Data Augmentation and Stain Color

Normalization in Convolutional Neural Networks for

Computational Pathology Medical Image Analysis,

Volume 58.

Titford, M. (2005). The long history of hematoxylin,

Biotechnic & Histochemistry, 80:2, pp. 73-78, Taylor

& Francis.

Zanjani F.G., Zinger S., Bejnordi, B.E., van der Laak,

J.A., de With, P.H. (2018). Stain normalization of

histopathology images using generative adversarial

networks, in: 2018 IEEE 15th International

Symposium on Biomedical Imaging, pp. 573–577,

IEEE.

Hematoxylin and Eosin Stained Histopathological Image Enhancement Method

385