Post-Processing of Thresholding or Deep Learning Methods for

Enhanced Tissue Segmentation of Whole-Slide Histopathological

Images

Michal Marczyk

1,2 a

, Agata Wrobel

, Julia Merta

and Joanna Polanska

Department of Data Science and Engineering, Silesian University of Technology, 44-100 Gliwice, Poland

Yale Cancer Center, Yale School of Medicine, 06511 New Haven, CT, U.S.A.

Faculty of Automatic Control, Electronics and Computer Science, Silesian University of Technology,

44-100 Gliwice, Poland

Keywords: Histopathology, Whole-Slide Image, Tissue Staining, Tissue Segmentation, Post-Processing.

Abstract: Digital pathology allows for the efficient storage and advanced computational analysis of stained

histopathological slides of various tissues. Tissue segmentation is a crucial first step of digital pathology

aimed at eliminating background, pen markings, and other artifacts, reducing image size, and increasing the

efficiency of further analysis. In most cases, color thresholding or deep learning models are used, but their

effectiveness is reduced due to complex artifacts and huge color variations between slides. We propose a post-

processing method to increase the tissue segmentation performance of any initial segmentation algorithm.

Using a set of 197 manually annotated histopathological images of breast cancer patients and 63 images of

endometrial cancer patients, we tested our method with 3 thresholding techniques and 3 deep learning-based

algorithms by calculating the Dice index, Jaccard index, precision, and recall. In both datasets, applying post-

processing increased precision and recall for thresholding methods and mostly precision for deep learning

models. Overall, applying post-processing gave better tissue segmentation performance than initial

segmentation methods, significantly increasing Dice and Jaccard indices. Our results proved that thanks to

post-processing, the tissue segmentation pipeline is more robust to noises and artifacts commonly present in

histopathological images.

1 INTRODUCTION

Modern medical imaging enables precise assessment

of the stained histopathologic slides of different

tissues. The most common staining method is

hematoxylin and eosin (HE) since using only these

two stains allows the visualization of major tissue

structures. Hematoxylin stains nuclei blue, while

eosin stains the cytoplasm and connective tissue pink

or red. With histopathological slides, pathologists can

determine the presence and stage of the disease or the

effect of medical treatment (Cooper, L.A., et al.,

2018). The development of digital scanners that can

obtain high-resolution whole-slide images (WSIs)

has contributed to creating extensive datasets with

images for various diseases. Digital processing of

slides and the decreasing data storage costs stimulated

https://orcid.org/0000-0003-2508-5736

https://orcid.org/0000-0001-8004-9864

the development of new algorithms in the fields of

image processing and machine learning (Serag, A., et

al., 2019). However, the main obstacle is that scanned

images are much larger than natural images, so

analysis could be extremely time-consuming with

high computational capacity requirements.

Most WSIs contain a lot of the background area,

which is uninformative for pathologists and only

increases the computational cost of image analysis.

Tissue segmentation leads to the accurate

identification and isolation of relevant regions of

WSI, which can significantly impact the efficiency

and speed of analysis and decrease the cost of data

storage (Salvi, M., et al., 2021). It is also crucial when

supervised methods are developed on digital slides

since it prevents learning from background noise.

Known semi-automated and manual segmentation

Marczyk, M., Wrobel, A., Merta, J. and Polanska, J.

Post-Processing of Thresholding or Deep Learning Methods for Enhanced Tissue Segmentation of Whole-Slide Histopathological Images.

DOI: 10.5220/0013174700003911

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

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 229-238

ISBN: 978-989-758-731-3; ISSN: 2184-4305

229

methods, such as color classification, edge detection,

and region growing, can be time-consuming and

labor-intensive, especially when dealing with WSIs.

Thus in practice one of the two approaches is chosen:

traditional thresholding with morphological

operations (Alomari, R.S., et al., 2009; Song, Y., et

al., 2023) or deep learning-based (DL-based)

methods (Riasatian, A., et al., 2020; Lucassen, R.T.,

et al., 2024). In DL models, an encoder is often used

to extract image features while a decoder is used to

restore extracted features to the original image size

and output the final segmentation results, like in U-

Net architecture (Riasatian, A., et al., 2020).

SlideSegmenter, based also on convolutional neural

network encoder-decoder architecture, introduces

some post-processing steps but only for dividing the

segmented tissue into cross-sections (Lucassen, R.T.,

et al., 2024). Additionally, SlideSegmenter provides

pen marking segmentation to exclude these regions

from tissue segmentation. Even though few

techniques have been introduced, there are still many

problems unsolved: (i) the requirement for the color

of the input image to be normalized, due to

differences in staining between laboratories; (ii)

manual adjustments of parameters for atypical cases;

(iii) the performance of the DL models depends on

the datasets used in developing the algorithm or

model training; (iv) supervised methods require

annotation by an expert pathologist.

Post-processing refinements like hole filling and

noise reduction could improve initial segmentation

performance. For example, traditional image

processing techniques produce initial segmentations

that contain single-pixel errors. Also, they might

work well on images with high contrast between

tissue and background but could struggle with images

containing a lot of noise or debris. WSIs often contain

artifacts, such as pen markings, air bubbles, and tissue

folds, that can interfere with subsequent analyses.

Post-processing techniques can be employed to

specifically detect and remove these artifacts,

resulting in cleaner and more reliable data for

downstream tasks. Lastly, DL models also often

benefit from post-processing steps that can refine

their predictions and improve their accuracy.

Here, we propose a method for post-processing

results of various tissue segmentation methods. We

tested a combination of our method with three

traditional thresholding algorithms and three DL-

based solutions on a set of manually annotated HE-

stained histopathological images. For thresholding,

we chose adaptive methods that are unsupervised,

parameter-free, and robust to changes in color

intensity distribution between slides.

2 MATERIALS AND METHODS

2.1 Data

Randomly selected 197 histopathological images of

breast cancer patients (BRCA) and 63 images of

uterine corpus endometrial cancer patients (UCEC)

from The Cancer Genome Atlas (TCGA) obtained

through The Cancer Imaging Archive (Clark, K., et

al., 2013) were manually annotated by a single expert.

The data were saved in .svs format that included

slides in different magnifications. For manual

annotation, images scanned with a magnification of

2.5x were selected. The annotation was done in a

MATLAB environment using the roipoly() function

or using ImageJ software. All fragments of the tissue

in a single slide were marked, excluding artifacts like

pen markings, shades, and others, and extremely

small fragments.

2.2 Adaptive Image Thresholding

Methods

Each scanned HE-stained image is composed of three

channels, R, G, and B, that represent color

components. Thresholding is done on each color

separately. Three thresholding methods were tested:

(i) GaMRed; (ii) Otsu; and (iii) Peaks. GaMRed is

based on the Gaussian mixture decomposition of 1D

signal and includes unique data cleaning and post-

processing steps (Marczyk, M., et al., 2020). Color

intensity from each channel was decomposed into 2

Gaussian components. The component with the

higher mean intensity represents background pixels.

The threshold value was found as the intersection

point between two Gaussians. The Otsu thresholding

method (Otsu, N., 1979) was implemented as a two-

step algorithm. First, the Otsu method is applied to

the original image color channels and cut-off values

are found. Second, the Otsu method is applied to

image color channels thresholded using cut-off values

from the first step, and new cut-off values are

estimated. The Peaks method is based on the peak

detection algorithm developed for the analysis of 1D

mass spectrometry data (Marczyk, M., et al.,2017).

The algorithm finds all maxima and minima using the

first derivative and then removes small amplitude

peaks, similar intensity peaks in close neighborhoods,

and the one with too small signal-to-noise ratio. For

each color, the threshold value was found as a

minimum between the two last peaks of color

intensity. All 3 methods allow adaptive estimation of

cut-of values for color thresholding without any

parameter tuning.

BIOIMAGING 2025 - 12th International Conference on Bioimaging

230

2.3 Deep Learning-Based Methods

Three existing methods based on deep learning (DL)

models were evaluated for a tissue segmentation task.

All models were previously trained by the respective

authors on their datasets and used as predictors with

TCGA data to mimic real life scenario in which tissue

annotations are not provided. All parameters were set

as default. The first method, called SlideSegmenter is

a convolutional neural network with a post-

processing method based on clustering predicted

centroid locations of the cross-sections in a 2D

histogram (Lucassen, R.T., et al., 2024). Another two

methods resulted from the experiments on the U-Net

architecture with different network backbones

(Riasatian, A., et al., 2020). Based on the published

results, the two best backbones were selected for

comparison: EfficientNet-B3 and MobileNet.

2.4 Post-Processing Methods

The proposed method consists of three subsequent

steps: (i) artifacts removal (P1); (ii) region filling

(P2); and (iii) small regions removal (P3). Artifacts,

which mainly are due to errors during specimen

preparation, staining, imaging, or tissue handling,

were found and removed using raw WSIs (without

thresholding). First, pixels including black and grey

color artifacts (resulting from tissue folding, air

bubbles, dust, debris and others) were identified using

the following steps: (i) create a mask of pixels with a

difference between red and green color channels

smaller than 10 and a difference between green and

blue color channels smaller than 10; (ii) remove the

background region resulting from image thresholding

from the mask; (iii) remove too small regions from

the mask (<30 pixels); (iv) apply morphological

closing using a disk of radius 3; (v) apply

morphological opening using a disk of radius 3.

Pixels including green color artifacts (mostly due to

green pen markings) were identified similarly, with

modifications only in the first step; the initial mask

was created using pixels with a difference between

red and blue color channels smaller than 10 and

intensity of green color higher than 150. Next, using

cleaned images an initial segmentation operation was

performed to find the tissue mask using methods

described in the previous section. Additionally, pixels

with low chroma component (square root of the sum

of squared a and b color values resulting from image

transformation to LAB color space smaller than 3)

were removed from the mask. Next, in P2 holes in the

mask were filled using the morphological region

filling method with 4 connectivity, and

morphological opening using a disk of radius 3 was

applied. Finally, in P3 small area objects were

removed from a tissue mask (smaller than 1% of the

total tissue mask region). All parameters were

estimated on a small pool of HE-stained images

scanned with 2.5x magnification, but not used in this

manuscript, and then fixed during the analysis. Codes

for thresholding and post-processing are available on

GitHub under the following adress:

github.com/ZAEDPolSl/WSI_TissueSeg .

2.5 Evaluation Metrics

Four different performance metrics were used to

compare tissue segmentation models before and after

applying post-processing: Dice coefficient, Jaccard

index, pixel precision, and pixel recall. These metrics

are defined as follows:

𝐷𝑖𝑐𝑒 =

2∗|𝑀𝑎𝑠𝑘∩𝐺𝑇|

𝑀𝑎𝑠𝑘

𝐺𝑇

(1)

𝐽

𝑎𝑐𝑐𝑎𝑟𝑑 =

|𝑀𝑎𝑠𝑘 ∩ 𝐺𝑇|

𝑀𝑎𝑠𝑘 ∪ 𝐺𝑇

(2)

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =

|𝑀𝑎𝑠𝑘 ∩ 𝐺𝑇|

𝑀𝑎𝑠𝑘

(3)

𝑅𝑒𝑐𝑎𝑙𝑙 =

|𝑀𝑎𝑠𝑘 ∩ 𝐺𝑇|

𝐺𝑇

(4)

In all equations, Mask represents all pixels within

the tissue mask, GT represents all pixels within

manual annotation (ground truth), and || represents

cardinality, which is a sum of pixels in the specified

area.

3 RESULTS AND DISCUSSION

We created a pipeline for the segmentation of tissue

regions on whole slide images, that is composed of

two subsequent steps (Figure 1A): (i) initial image

segmentation to eliminate the background area of the

tissue slide; (ii) image and tissue mask post-

processing to refine initial segmentation and isolate

only relevant tissue fragments. We tested different

methods of thresholding, which estimate the

background cut-off values based on the analysis of

histograms of color intensities and DL-based models.

For the resulting images, we applied image

processing techniques used in computer vision which

are necessary to increase the quality of the tissue

segmentation by removing artifacts, and all other

noise components of the image (see example in

Figure 1B).

Post-Processing of Thresholding or Deep Learning Methods for Enhanced Tissue Segmentation of Whole-Slide Histopathological Images

231

Figure 1: Proposed tissue segmentation algorithm including post-processing step. A) Subsequent steps of the full pipeline. B)

Exemplary results of applying the algorithm to the HE image: from raw image (left), through initial tissue segmentation

(middle) to segmented tissue after post-processing step (right).

Figure 2: Comparison of tissue segmentation performance between initial segmentation methods (without post-processing)

using Dice and Jaccard indices for BRCA (A) and UCEC (B) cohorts.

3.1 Initial Segmentation of WSIs

We compared three thresholding methods with three

different DL-based methods using Dice and Jaccard

indices and precision and recall metrics (Figure 2).

Overall, DL-based methods gave better results than

traditional methods on both datasets (Table 1).

Among thresholding methods, we observed the

highest values for GaMRed (median Dice = 0.9556 in

BRCA and 0.9725 in UCEC; median Jaccard =

0.9149 in BRCA and 0.9464 in UCEC) while the

lowest for the Peaks method. Among DL-based

method, we observed the highest values for

MobileNet in BRCA (median Dice = 0.9749; median

Jaccard = 0.9511) and EfficientNet-B3 in UCEC

(median Dice = 0.9913; median Jaccard = 0.9827).

Also, DL-based methods gave higher minimum

values of indices than thresholding methods,

especially in UCEC (Figure 2B).

BIOIMAGING 2025 - 12th International Conference on Bioimaging

232

GaMRed

Otsu

Peaks

EfficientNet-

Mobile

Net

Slide

Segmenter

Figure 3: Selected cases with poor initial segmentation (no post-processing). GT is a ground truth segmentation (yellow line

around tissue). For each method, the green color indicates true positive, blue false negative, and red false positive regions.

On average, thresholding methods gave higher

precision than DL-based methods, but lower recall

(Table 1). We found the best precision for Peaks

(median = 0.9793 in BRCA and 0.9976 in UCEC)

while the worst for SlideSegmenter (median = 0.9461

in BRCA and 0.9933 in UCEC). On the opposite, the

best recall was for Slide Segmenter in BRCA (median

= 0.9970) and EfficientNet-B3 in UCEC (median =

0.9902). The worst recall was for the Peaks method

(median = 0.9193 in BRCA and 0.9437 in UCEC).

In Figure 3, we visualized the worst-case WSI for

thresholding methods (left), and DL-based methods

(right) in terms of Jaccard and Dice indices. For

thresholding methods, we observed mostly false

negative regions (colored blue), where the

segmentation algorithm omitted tissue regions

Post-Processing of Thresholding or Deep Learning Methods for Enhanced Tissue Segmentation of Whole-Slide Histopathological Images

233

selected by the expert giving lower recall. However,

in the original images, excluded regions were mostly

composed of adipose tissue, which is not relevant in

many digital pathology tasks. For the DL-based

method, we observed mostly false positive regions

(colored red), where the segmentation algorithm

marked a larger area than the expert giving lower

precision.

A closer investigation of ground truth images

shows that areas with noise or artifacts were wrongly

selected as tissue regions by DL-based methods,

which might bring worse consequences in further

analysis.

3.2 Influence of Post-Processing on

Tissue Segmentation Results

For both datasets and each initial segmentation

method, we found a significant increase in almost all

performance metrics (p-value from Wilcoxon test

smaller than 0.05) after applying the post-processing

step (Figure 4 and Tables 1 and 2). Only for

EfficientNet-B3 and Slide Segmenter, there was a

small decrease in recall in the BRCA dataset (median

equal to -0.0001 and -0.0003, respectively).

In terms of Dice and Jaccard indices, a higher gain

was observed for thresholding methods than DL-

based methods in both datasets, which leads to a

similar final performance for all methods (Table 1).

Among unsupervised methods, after post-processing,

we observed the highest values for GaMRed (median

Dice = 0.9865 in BRCA and 0.9941 in UCEC; median

Jaccard = 0.9733 in BRCA and 0.9883 in UCEC)

while the lowest for the Peaks method. Among DL-

based methods, the results are similar giving U-Net-

based methods better than Slide Segmenter. Even

after post-processing, thresholding methods gave

higher precision and lower recall than DL-based

methods, but the difference between them is now

smaller. Again, we found the best precision for Peaks

(median = 0.9923 in BRCA and 0.997 in UCEC)

while the worst for Slide Segmenter (median = 0.9657

in BRCA and 0.9964 in UCEC). The best recall was

found for Slide Segmenter in BRCA (median =

0.9971) and EfficientNet-B3 in UCEC (median =

0.9976) and the worst for the Peaks method (median

= 0.9836 in BRCA and 0.9881 in UCEC).

In Figures 5 and 6, we visualized selected WSI,

for which the number of false positive and false

negative pixels is significantly reduced after post-

processing for each initial segmentation method. The

selected image from BRCA (Figure 5) contains a lot

of small holes in the tissue, which caused poor

segmentation with thresholding methods. The

selected image from UCEC (Figure 6) contains a blue

pen marking, that was segmented as tissue mostly for

DL-based methods. Also, grey shades below the

tissue were falsely marked by Slide Segmenter. All

these mistakes were reduced by applying proposed

post-processing method.

Figure 4: Gain in tissue segmentation performance of

different initial segmentation methods after applying post-

processing measured using Dice and Jaccard indices for

BRCA (A) and UCEC (B) cohorts.

BIOIMAGING 2025 - 12th International Conference on Bioimaging

234

No post-processing After post-processing

GaMRed

Otsu

Peaks

EfficientNet

-B3

Mobile

Net

Slide

Segmenter

Figure 5: Selected case from BRCA dataset with high gain in performance after post-processing (right). . For each method,

the green color indicates true positive, blue false negative, and red false positive regions.

Post-Processing of Thresholding or Deep Learning Methods for Enhanced Tissue Segmentation of Whole-Slide Histopathological Images

235

No post-processing After post-processing

GaMRed

Otsu

Peaks

EfficientNet-

Mobile

Net

Slide

Segmenter

Figure 6: Selected case from UCEC dataset with high gain in performance after post-processing (right). For each method, the

green color indicates true positive, blue false negative, and red false positive regions.

BIOIMAGING 2025 - 12th International Conference on Bioimaging

236

Table 1: Median values of performance indices across all images for each dataset and initial segmentation method. Results

before and after post-processing are shown. Bold highlights the best value of each index within the dataset and pipeline step.

Dataset Step Method Dice Jaccard Precision Recall

BRCA

Initial

segmentation

GaMRe

0.9556 0.9149 0.9736 0.9449

Otsu 0.9528 0.9099 0.9759 0.9316

Peaks 0.9498 0.9043 0.9793 0.9193

EfficientNet-B3 0.9719 0.9454 0.9511 0.9961

MobileNet 0.9749 0.9511 0.9604 0.9936

SlideSegmente

0.9692 0.9403 0.9461 0.9970

Post-processed

GaMRe

0.9865 0.9733 0.9901 0.9862

Otsu 0.9859 0.9723 0.9910 0.9848

Peaks 0.9857 0.9719 0.9923 0.9836

EfficientNet-B3 0.9833 0.9671 0.9711 0.9965

MobileNet 0.9856 0.9715 0.9767 0.9955

SlideSegmente

0.9798 0.9603 0.9657 0.9971

UCEC

Initial

segmentation

GaMRe

0.9725 0.9464 0.9974 0.9487

Otsu 0.9697 0.9411 0.9974 0.9437

Peaks 0.9681 0.9382 0.9976 0.9437

EfficientNet-B3 0.9913 0.9827 0.9935 0.9902

MobileNet 0.9905 0.9812 0.9950 0.9870

SlideSegmente

0.9898 0.9798 0.9933 0.9893

Post-processed

GaMRe

0.9941 0.9883 0.9997 0.9891

Otsu 0.9936 0.9872 0.9997 0.9885

Peaks 0.9938 0.9878 0.9997 0.9881

EfficientNet-B3 0.9967 0.9935 0.9968 0.9976

MobileNet 0.9970 0.9940 0.9977 0.9961

SlideSegmente

0.9966 0.9932 0.9964 0.9975

Table 2: Gain in performance indices across all images for each dataset and initial segmentation method. Results show the

median difference of results after post-processing and initial segmentation. Bold highlights the best value of each index within

the dataset.

Dataset Method Dice Jaccard Precision Recall

BRCA

GaMRe

0.0281 0.0532 0.0121 0.0389

Otsu 0.0322 0.0601 0.0105 0.0476

Peaks 0.0341 0.0632 0.0097 0.0541

EfficientNet-B3 0.0088 0.0171 0.0158 -0.0001

MobileNet 0.0076 0.0148 0.0115 0.0006

SlideSe

mente

0.0091 0.0171 0.0164 -0.0003

UCEC

GaMRe

0.0183 0.0357 0.0022 0.0333

Otsu 0.0206 0.0400 0.0022 0.0386

Peaks 0.0209 0.0402 0.0020 0.0385

EfficientNet-B3 0.0052 0.0102 0.0025 0.0080

MobileNet 0.0059 0.0117 0.0021 0.0095

SlideSe

mente

0.0061 0.0121 0.0027 0.0083

3.3 Ablation Study

Lastly, we tested the influence of subsequent post-

processing methods on the performance of the

proposed tissue segmentation pipeline. For each

initial segmentation algorithm, we observed similar

patterns of changes in the Dice and Jaccard indices

(Figure 7). The highest increase was obtained after

artifact removal together with region filling (IS+P12;

3% on average), and then a small increase was found

after filtering too small regions (IS+P123) in BRCA,

but not in UCEC, where only DL-based methods

showed a small increase. Artifact removal alone had

the smallest influence on tissue segmentation

performance. For DL-based methods, we even

observed a small decrease in the Dice index mostly in

UCEC. However, we noticed that without this step

region filling and small regions removal steps gave

much worse results.

Post-Processing of Thresholding or Deep Learning Methods for Enhanced Tissue Segmentation of Whole-Slide Histopathological Images

237

Figure 7: Tissue segmentation results after subsequent steps

of the tissue segmentation in BRCA (A) and UCEC (B). IS

represents the initial segmentation step, while P123 artifacts

removal, region filling, and small regions removal.

4 CONCLUSIONS

Segmentation of tissue regions on whole-slide images

is an important first step in the advanced computational

analysis of stained histopathological slides. We

developed a post-processing algorithm that was

successfully applied to simple image thresholding

methods and more advanced DL-based models. Our

analysis proved that the proposed tissue segmentation

pipeline is robust to noise and different artifacts

observed in the sample, and it can consistently acquire

better results than initial segmentation alone.

Regardless of a small improvement in performance

indices, we visualized some cases to provide visual

proof of post-processing necessity. Lastly, all

parameters of the proposed method were selected on

other, unseen data (but scanned with the same

magnification), and fixed during analysis. Thus, there

is a potential to improve the results even more through

the parameter tuning procedure.

ACKNOWLEDGMENTS

This publication was supported by the Excellence

Initiative - Research University program imple-

mented at the Silesian University of Technology,

grant no. 02/070/SDU/10-21-02 (MM), COMPASS-

NMD Project funded by the European Union Horizon

Europe program under Grant Agreement 101080874

(JP) and Silesian University of Technology grant no.

02/070/BK_24/0052 for maintaining and developing

research potential.

REFERENCES

Cooper, L.A., Demicco, E.G., Saltz, J.H., Powell, R.T., Rao,

A., Lazar, A.J.: PanCancer insights from The Cancer

Genome Atlas: the pathologist's perspective. The Journal

of pathology 244, 512-524 (2018)

Serag, A., Ion-Margineanu, A., Qureshi, H., McMillan, R.,

Saint Martin, M.-J., Diamond, J., O'Reilly, P., Hamilton,

P.: Translational AI and deep learning in diagnostic

pathology. Frontiers in medicine 6, 185 (2019)

Salvi, M., Acharya, U.R., Molinari, F., Meiburger, K.M.: The

impact of pre-and post-image processing techniques on

deep learning frameworks: A comprehensive review for

digital pathology image analysis. Computers in Biology

and Medicine 128, 104129 (2021)

Alomari, R.S., Allen, R., Sabata, B., Chaudhary, V.:

Localization of tissues in high-resolution digital

anatomic pathology images. In: Medical Imaging 2009:

Computer-Aided Diagnosis, pp. 349-358. SPIE

Song, Y., Cisternino, F., Mekke, J.M., de Borst, G.J., de

Kleijn, D.P., Pasterkamp, G., Vink, A., Glastonbury,

C.A., van der Laan, S.W., Miller, C.L.: An automatic

entropy method to efficiently mask histology whole-slide

images. Scientific Reports 13, 4321 (2023)

Riasatian, A., Rasoolijaberi, M., Babaei, M., Tizhoosh, H.R.:

A comparative study of U-net topologies for background

removal in histopathology images. In: 2020 International

Joint Conference on Neural Networks (IJCNN), pp. 1-8.

IEEE

Lucassen, R.T., Blokx, W.A., Veta, M.: Tissue Cross-Section

and Pen Marking Segmentation in Whole Slide Images.

arXiv preprint arXiv:2401.13511 (2024)

Clark, K., Vendt, B., Smith, K., Freymann, J., Kirby, J.,

Koppel, P., Moore, S., Phil-lips, S., Maffitt, D., Pringle,

M., Tarbox, L., Prior, F.: The Cancer Imaging Archive

(TCIA): Maintaining and Operating a Public Information

Repository. Journal of Digital Imaging 26, 1045-1057

(2013)

Marczyk, M., Jaksik, R., Polanski, A., Polanska, J.:

GaMRed—Adaptive Filtering of High-Throughput

Biological Data. IEEE/ACM Transactions on

Computational Biology and Bioinformatics 17, 149-157

(2020)

Otsu, N.: A Threshold Selection Method from Gray-Level

Histograms. IEEE Transactions on Systems, Man, and

Cybernetics 9, 62-66 (1979)

Marczyk, M., Polanska, J., Polanski, A.: Improving peak

detection by Gaussian mixture modeling of mass spectral

signal. In: 2017 3rd International Conference on

Frontiers of Signal Processing (ICFSP), pp. 39-43. IEEE

BIOIMAGING 2025 - 12th International Conference on Bioimaging

238