Evaluating the Impact of Color Normalization on Kidney Image Segmentation

International Journal on Cybernetics & Informatics (IJCI) Vol. 12, No.5, October 2023
Dhinaharan Nagamalai (Eds): EMVL, EDUT, SECURA, AIIoT, CSSE -2023
pp. 93-105, 2023. IJCI – 2023 DOI:10.5121/ijci.2023.120509
EVALUATING THE IMPACT OF COLOR
NORMALIZATION ON KIDNEY
IMAGE SEGMENTATION
Sai Javvadi
University of Louisville, Louisville Kentucky, USA
ABSTRACT
The role of deep learning in the recognition of morphological structures in histopathological data has
progressed significantly. But, less intensive preprocessing stages and their contribution to deep learning
pipelines is often overlooked. Color normalization (CN) algorithms are among the most prominent methods
in this stage, and they work by standardizing the staining pattern of a dataset. However, the impact of
various color normalization algorithms on the detection of glomeruli functional tissue units (FTUs) in
kidney tissue data has not been explored before. An advanced deep learning architecture was built with the
U-NET segmentation model. The U-NET model is an architecture that specializes in the segmentation of
biomedical data. A dataset of 15 kidney whole slide images (WSIs), each annotated with locations of
glomeruli FTUs were processed and subsequently normalized according to three 3 different conventional
color normalization techniques (Reinhard, Vahadane, Macenko), and fed into a U-NET model. The dice
score coefficient (DSC) was used to compare the results of each run. It was determined that color
normalization algorithms significantly impact the segmentation results of deep learning algorithms, with
the Reinhard algorithm being the best technique. The implications of this work are immense, as it could
contribute to the proliferation of color normalization techniques in preprocessing deep learning workflows,
which could improve general segmentation accuracies.
KEYWORDS
Deep Learning, Color Normalization, Histopathology, Kidney, Glomeruli
1. INTRODUCTION
1.1. Overview
The application of deep learning within the context of medical imaging has allowed for faster and
more accurate analysis compared to conventional methods of analysis that rely on a pathologist.
The advantages that deep learning provides in this context can be highlighted when dealing with
organ tissues that exhibit great heterogeneity, such as the kidney. The great diversity of tissue
within the kidney makes it especially difficult for pathologists to annotate, consequently making
deep learning techniques focused on the annotation of kidney tissue more valuable and significant
[1]. Thus, exploring ways to make deep learning models trained on histopathological kidney
tissue data more accurate and useable within a clinical setting is imperative. By being able to
effectively evaluate particular techniques that could contribute to the enhancement of deep
learning model performance and applicability when applied to kidney tissue data, similar strides
could be taken with other kinds of tissue data. The widespread application of deep learning in
pathological contexts, and thus for the diagnosis and prognosis of multi-organ diseases, could
revolutionize the medical sphere [2].

94
Figure 1. Histopathological Image Segmentation Workflow
The purpose of this research is to evaluate the impact of various, conventionalized color
normalization techniques on a self-engineered U-NET deep learning model focused on analyzing
kidney tissue data. This would allow for the changes in model performance according to color
normalization technique to be quantified, and could allow for the greater application of particular
color normalization techniques in more contexts. This could also effectively improve model
performance and make strides in achieving clinical integration.
Figure 2. Color Normalization Workflow – Pattern of reference image applied to diverse source dataset for
normalized generated dataset
1.2. Literature Review
Previously developed deep learning models trained to identify glomerular functional tissue units
have often emphasized the performance of the U-NET model for its high accuracy [3, 4]. The U-
NET model was proposed in 2015, and is a deep learning model especially compatible with
biomedical imaging due to its unique architecture [5]. Proposed improvements in previously
referenced works are the inclusion of a preprocessing technique that normalized stains in the
datasets [3]. Further work involves the development of a computer-aided diagnostic (CAD)
model that is capable of not only segmenting, but also classifying glomeruli [6]. Although these
works demonstrate a variation in methodology in terms of the model leveraged (an ANN rather

95
than the more commonly used CNN), an emphasis on the experimentation of various
preprocessing techniques, namely color normalization, is generally absent. Previous work that
has focused on the application of various color normalization algorithms and their impact on deep
learning models trained on different kinds of histopathological data [6] allows for the techniques
that are potentially effective with kidney tissue data to be narrowed down, despite this work not
experimenting with this type of tissue in particular. Other work that focused on gauging the
impact of color normalization techniques on other kinds of biomedical data have found that it
successfully tamed variability in data [7].
The lack of research surrounding the application and evaluation of color normalization
techniques on kidney histopathological images is a clear research gap. This project aims to
address this gap by testing 3 color normalization techniques that are most commonly used in
order to determine the impact of color normalization on deep learning model performance,
potentially allowing for heightened model performance and greater clinical applicability. This
project hypothesizes that every single normalization technique will aid in a model that improves
upon the performance of a deep learning model trained solely on the original, unnormalized data.
Furthermore, of the 3 color normalization algorithms being tested – Vahadane, Reinhard, and
Macenko – it is hypothesized that the Reinhard algorithm will perform the best due to its frequent
use in literature.
1.3. U-NET Model Overview
The U-NET model is able to generally provide better segmentation accuracies when applied to
histopathological data due to its particular architecture, as seen in Figure 6. The first half of its U-
shape is known as the contracting path, which primarily consists of convolutional and max
pooling layers. These layers are used to extract features, and to reduce the dimensionality of
feature maps. The use of these layers allowing for complex and high-level features to be
captured. The second half of the model is composed of the expanding path. This path creates the
segmentation map and leverages transposed convolution (or Up-convolution) to increase the
resolution of the feature maps. Essentially, this path transforms the compressed representation
from the contracting path to a segmentation map in the original resolution. One of the most
unique aspects of the U-NET architecture are the skip connection layers (lines connecting the
contracting path to the expanding path). These connections allow for the preservation of very
specific spatial information. This is especially beneficial when applied to extremely complex and
detailed data like that of histopathological data. Consequently, when applied to this kind of data,
more accurate and precise segmentations can be achieved.
2. METHEDOLOGY
Kidney tissue data is derived from the “HuBMAP: Hacking the Kidney Dataset”. This dataset
presents 20 total Formalin Fixed Paraffin Embedded (FFPE) Periodic Acid-Schiff (PAS)-stained
kidney whole slide images (WSIs). The data is pre-annotated with the locations of Glomeruli
Functional Tissue Units (FTUs). From each image, individual tiles of resolution 512 x 512 are
obtained; this processing stage results in roughly 2600 images. In order to proliferate the data,
augmentative procedures are used: random cropping, random mirroring, random jittering, and
random noise. Examples of these augmentation on data is demonstrated in Figure 4. These
augmentative procedures create transformation within the data so as to create more variations and
to expand the size of the dataset. Furthermore, augmentations contribute to a more generalized
and accurate deep learning model. This proliferated dataset is then portioned for training/testing
purposes with a 90/10 ratio (90% of the dataset is allocated for training, while the other 10% is
set aside for testing).

96
Figure 3. Example of Kidney Data - Tile (left) and Annotated Tile (right)
Figure 4. Augmentative Functions - Left: Original Image; Right: Augmented Image
A commonly selected reference image is chosen based upon the effectiveness of the reference
image established by successful teams who had previously worked with the dataset. The original
dataset is normalized based upon a singular reference image by each of the three-color
normalization algorithms. In total, there are 4 datasets (the original, and 3 normalized). An array
of color metrics is applied to each dataset to gauge the impact of color normalization on the
quality of the dataset. Subsequently, each dataset is fed into a U-NET deep learning model
trained to recognize glomeruli FTUs. The dice score coefficient (DSC) is used to evaluate the
performance of the model for each dataset, effectively comparing the impact of color
normalization on model accuracy. These results leverage 10-fold cross validation to ensure
statistical significance.
Figure 5. Applying Reference Pattern to Source Dataset – A generated dataset more aligned in color is
generated by applying the pattern of the reference image to the original dataset

97
2.1. Color Normalization
Color Normalization (CN) algorithms were specifically chosen for experimentation due to their
prior success in other tissue-based segmentation algorithms [8]. The successful use of the
Reinhard, Vahadane, and Macenko CN algorithms for medical segmentation tasks may be
potentially translated to glomeruli detection algorithms.
Reinhard et al developed a globalized technique for color normalization, in which the source
image receives the mean color values of the target, or reference image. This results in a generated
image that is very similar to that of the reference [9].
Macenko and Vahadane developed color normalization techniques that leveraged stain
separation. These separated stain maps are normalized and combined in a fashion that relies on
the stain color of the reference image [10].
2.2. Architecture
A constant architecture must be used because with this constant, an accurate measure of the
impact of CN algorithms would not be available. Furthermore, studies that test various data
augmentations keep their models constant as changing the models can impact the overall
accuracy, interfering with the comparison of different preprocessing algorithms and creating
distorted comparisons. For example, previous works that evaluate the strength of color
normalization methods have kept particular model architectures as constants in their experiments
[20]. In this particular project, the U-NET model architecture is specifically used is due to the
prevalence of its usage in various medical segmentation purposes [11]. The architecture is
specialized to work with medical data that is meant to be segmented, and operates in a fast and
precise manner [12]. In the context of identifying glomeruli specifically, recent literature has
widely used the U-NET algorithm [13]. Therefore, using U-NET as a constant architecture
ensures a good baseline performance while practically measuring any added benefit of a added
CN algorithm.
Figure 6. U-NET Deep Learning Model Architecture – Retrieved from [14]

98
The particular U-NET model architecture relies on the Dice Score Loss function, which
quantifies the difference in segmentation between a model’s predicted segmentation and the
ground truth. Furthermore, the Adam optimizer is leveraged at learning rate 1e-4 in order to
adjust the weights of the deep learning model so as to achieve a set of weights capable of optimal
model performance. Furthermore, the model is trained on 100 epochs with a batch size of 8. This
means that the model goes through the entirety of the training data 100 times, and is updated
whenever it processed 8 images.
To reiterate, the novelty of this work is derived from the direct comparison of model performance
when trained on datasets processed with various color normalization techniques (Reinhard,
Vahadane, & Macenko). In the particular context of segmenting Glomeruli FTUs in kidney
images, there is a lack of this work. Demonstrating the effectiveness of these processing
techniques with morphologically complex data may contribute to its implementation in wider
scales, and result in residual benefits such as mitigating bias in deep learning models (bias
accumulates from heavy diversity in stains).
2.3. Performance Metrics
Once the deep learning model is developed, performance metrics need to be compared to
determine whether models with CN preprocessing stages present any added benefits. The usage
of performance metrics has long characterized the effectiveness of machine learning algorithms
[15]. Therefore, using the correct performance metrics is vital. For segmentation, prevalent
performance metrics include the Dice Score Coefficient (DSC), IoU (Intersection-over-Union),
and accuracy. Previous works have used the DSC and IoU metrics to directly compare and
evaluate various models trained of kidney-based glomerular data [16]. Consequently, the use of
these metrics as a ground for comparison among a variety of models suggests that it is a reputable
method for the accurate reflection of model performance. For this research, the DSC was adopted
as a means to quantify model performance.
Equation 1. Dice Coefficient
The DSC value measures the intersection between two areas. It is 2 * (area of intersection) / total
number of pixels in two areas. The greater this value, the more indicative of a model’s strength,
as the predicted segmentation is closer to the ground truth (real) segmentation.
The use of PAS-stained data to feed into U-NET algorithms is commonplace in this field.
Therefore, comparing this workflow to one that uses a color normalization algorithm will provide
valuable insight into the latter’s impact on model performance while also being practical. This
methodology will allow the impact of color normalization algorithms to properly be evaluated
and potentially make large strides in the use of CN algorithms as a general preprocessing
implementation. There were two primary phases for testing the impact of color normalization
with glomeruli and histopathological data. The first step was to choose three state-of-the-art color
normalization methods: Reinhard, Macenko, and Vahadane. In order to compare the 3 different
methods of color normalization tested (the 3 experimental variables being the color normalization
techniques being tested, and the control being the baseline workflow without normalization), it
was imperative that an array of color metrics, or measurements that indicate particular qualities of

99
an image’s color nature, was used. Rather than use a singular color metric to compare the
performances of these different methodologies, an array of metrics was adopted due to the fact
that since color can be measured in a myriad of ways: intensity, hue, saturation, etc. The use of a
singular metric would not be indicative of the overall performance of a normalization technique
[1]. The particular metrics adopted were based on previous work that evaluated the impact of
color normalization through these metrics [17].
The particular metrics adopted include FSSIM, or the functional similarly index measure which
screens for the quality of a new image in relation to an original; UQI, or universal quality index,
which reflects the overall quality of an image; PCC, or the Pearson Correlation Coefficient,
which measures the strength of the relationship between two images.
Figure 7. First Testing Phase Workflow – The original dataset is normalized with each CN algorithm. 3
different color metrics (PCC, FSSIM, and UQI) are subsequently applied to each dataset in order to gauge
for quality.
Figure 8. Second Testing Phase Workflow – The normalized datasets, along with the original, are
subsequently fed into a deep learning model. Model performance is evaluated with the DSC metric.
2.4. Model Training Environment
With this methodology, various color normalization algorithms can be properly compared in
multiple facets. Through the application of different color metrics, the quality of the data they
produce can be evaluated. More importantly, the development of a U-NET deep learning model
allows for the implications of the differently normalized datasets on model performance to be
properly evaluated. The development of the aforementioned architectures will be conducted
natively using the Google Collaboratory service. By utilizing cloud-based high-end GPUs, the
training and testing requirements of this project may be fulfilled significantly faster.

100
3. RESULTS
Figure 9. Example of Normalized Images – From left to right: Original Image, Macenko Normalized,
Vahadane Normalized, Reinhard Normalized
3.1. Color Metric Performance Results
Comparing the qualities of the color normalized datasets in Figure 10, Vahadane possesses the
composite highest quality. While the Reinhard-normalized dataset possesses the highest average
PCC score, the Vahadane-normalized data boasts the highest average score for the FSSIM and
UQI metrics. Since both the FSSIM and PCC metrics evaluate the quality of an image in relation
to another, these metrics offer conflicting narratives about the quality of the generated images in
relation in to the original. However, the UQI metric analyzes general images quality and does not
do so in relation to another image. Based upon these results, a general ranking of the quality of
each of the normalized datasets can be assigned (in highest to lowest order): Vahadane, Reinhard,
Macenko.
Figure 10. Color Metric Array Comparison - Blue: PCC; Orange: FSSIM; Gray: UQI
From this initial analysis of the normalized data, it was deducted that since the Vahadane-
normalized data reflected attributes that suggested it was of higher quality than the other
normalized datasets, that there would potentially be a correlation to better performance. This is
because higher quality data may be more meaningful and beneficial for feature learning,
contributing to better model performances.

101
The quality and effect of normalization on data can be further visualized by comparing the nature
of the normalized data to that of the un-normalized, original dataset. The images were compared
by using the average Hue and Saturation channel ratios of every single image in the dataset. This
was done by converting each RGB image in all datasets into the HSV (Hue Saturation Value) file
format. Then, the hue and saturation channels were extracted and compared for each image. This
method is reflective of the effectiveness of color normalization, particularly within
histopathological images with a bimodal nature, or of a nature in which two hues dominate the
image when visualized with the hue channel [18].
Figure 11. Average Hue vs. Saturation Intensities for Non-Normalized and Normalized Datasets
As visualized above, every single normalization technique was able to effectively normalize the
original data. The average Saturation vs. Hue intensity is significantly more concentrated for each
of the normalized data ploys than that of the non-normalized data, which is indicative of effective
color normalization. Furthermore, the styles and natures of the normalization techniques can also
be observed with these plots. The Macenko and Vahadane plots are very similar, with a cluster of
points congregated in a vertically linear fashion towards the right-most portions of the plot. In
contrast, the Reinhard plot depicts a cluster of points arranged in a circular fashion, still closer to
the right-edge of the plot. This is expected since the Macenko and Vahadane techniques operate
very similarly, while the Reinhard method takes a very different approach in color
transformation.
Analyzing the plots of the normalized datasets, the quantity of outliers can be observed. Both the
Vahadane and Macenko normalized datasets seem to have a limited numbers of points that are
distant from a central cluster. In contrast, the Reinhard normalized dataset seems to have a
significant number of outlier points. Many points on the Reinhard plot are very distant from the
central cluster. The degree of alignment, although strong with all normalized datasets, appears to
be weaker in the Reinhard normalized dataset. This could potentially be indicative of less
effective normalization.

102
3.2. Model Segmentation Results
It is critical that the effects of these color normalization techniques on the segmentation of the
histopathological dataset be effectively tested. Since color normalization effects the nature of an
image, and deep learning models segment these images to precisely identify particular
morphological features, the effect of these techniques on deep learning segmentation was critical
for a thorough understanding of the practicality that color normalization offers in the clinical
context.
Therefore, it was important to test for the effectiveness of color normalization techniques on the
identification of glomerular structures within kidney WSIs.
For the statistical validation of this data K-folds cross validation (CV) was used; K-folds CV is a
commonly used tool to ensure that the results of machine learning performances are not merely
coincidental [19]. As demonstrated in Figure 10, the testing partition of the data is varied with
each run. With 10-fold CV, each dataset (original, and 3 normalized) is fed into the model 10
times, each time with a different validation set. Each run returned a value for the DSC, which
reflects a generalized performance of the model’s predicted segmentation accuracy.
Figure 12. K-Fold Cross Validation – Retrieved from [20]
Table 1. Average DSC Performances – Averages the DSC score outputted from a model trained on each
dataset
Average DSC with 10-Fold CV
Dataset Original Macenko Vahadane Reinhard
HuBMAP 0.7963 0.8712 0.8124 0.8842
As observed in Table 1, the Reinhard algorithm boasts the highest average DSC value. As this
result is the culmination of 10 different runs, each with different folds of training and test data,
this result is treated as non-coincidental and statistically-validated.
These results suggest that although every single normalization technique improved upon the
baseline performance of the deep learning model, the Reinhard algorithm was most successful in
doing so. The Reinhard algorithm was able to achieve an average DSC of 0.8842, a massive leap

103
in performance compared to the average DSC of 0.7963 with the original dataset. Furthermore,
the color metrics used to test the general quality of the produced images indicated that the
Vahadane technique produced better quality images, but despite this, did not outperform the
Reinhard technique. This result reinforces the prominence of the Reinhard algorithm in deep
learning workflows that do utilize color normalization.
4. CONCLUSION
To conclude, it was found that color normalization has a significant impact on the identification
of glomeruli within PAS-stained kidney whole slide images. Color normalization involves
standardizing stain variations that may exist within a dataset. Of the 3 common color
normalization algorithms tested, it was determined that all three of them resulted in higher
segmentation performance for the tested dataset, as according to the Dice Score Coefficient, than
the original dataset. This means that when fed with the normalized data, deep learning models,
regardless of what normalized data it was fed, was able to recognize more glomeruli with higher
precision than what it was able to when fed the original, non-normalized data. The quality of the
segmentations produced also differed drastically. Often times, the segmentation produced by the
model that was fed some type of normalized data was significantly more precise in a variety of
aspects. Greater precision can be exhibited by the sheer number of structures within an image that
was able to be identified by a deep learning model, and the edges of the segmentations (precise
segmentations have rougher edges while less precise segmentation are rounder and more
connected together, representing less precision). In order to obtain these results, it was imperative
that the deep learning model used for experimentation was kept constant. Interestingly enough, it
was determined that the original data was generally of a higher quality according to a myriad of
color metrics applied. This supports the idea that quality does not translate to segmentation
performance.
The implications of this research are strong – as more attention is paid to the application of
certain CN algorithms on the segmentation of not only kidney tissue, but other types of
histopathological images, the biases that are often inevitable is histopathological datasets may be
combatted. These are the same biases that render many, even highly-accurate models, impractical
within a hospital setting. This research hopes to support a narrative, that, color normalization can
aid in the reduction of bias present in stain data and make deep learning algorithms more suitable
and safer to use in a clinical setting.
4.1. Limitations
This research is limited by only sourcing data from a singular dataset. By agglomerating data
from diverse sources, the results of this study and corresponding application can be reinforced.
Furthermore, the size of the dataset itself was quite small (~2600 images without augmentative
procedures). In future studies, similar data from other datasets can be sourced and combined for
larger datasets. Furthermore, this study is limited by the use of only a singular deep learning
model (U-NET model) being used to test the effectiveness of normalization. Since only a single
deep learning model is used, the impact of different normalization techniques is not fully
portrayed. In future studies, testing multiple architectures will be considered so that more
representative results can be conveyed.

104
4.2. Future Direction
For future research, the application of new color normalization techniques can be explored and
compared to the effectiveness of non-normalized, original data and even compared against
conventional, state-of-the-art color normalization techniques. In this work, conventional
techniques were tested. However, with the breakthroughs of new generative models, such as
generative adversarial networks (GANs) in the field, new color normalization techniques are
being developed which could have significant implications for the landscape. Furthermore, more
model architectures other that U-NET can be tested.
REFERENCES
[1] Huo, Y., Deng, R., Liu, Q., Fogo, A. B., & Yang, H. (2021). AI applications in renal pathology.
Kidney International, 99(6), 1309–1320. https://doi.org/10.1016/j.kint.2021.01.015
[2] Furlow, B. (2017). Deep learning poised to revolutionise diagnostic imaging. The Lancet
Respiratory Medicine, 5(10), 779. https://doi.org/10.1016/s2213-2600(17)30292-8
[3] Shubham, Shubham, et al. “Identify Glomeruli in Human Kidney Tissue Images Using a Deep
Learning Approach.” Soft Computing, vol. 27, no. 5, 23 Aug. 2021, pp. 2705–2716,
https://doi.org/10.1007/s00500-021-06143-z.
[4] Wilbur, D. C., Smith, M. L., Cornell, L. D., Andryushin, A., & Pettus, J. R. (2021). Automated
identification of glomeruli and synchronized review of special stains in renal biopsies by Machine
Learning and Slide Registration: A cross‐institutional study. Histopathology, 79(4), 499–508.
https://doi.org/10.1111/his.14376
[5] Piórkowski, A., & Gertych, A. (2018). Color normalization approach to adjust nuclei segmentation
in images of hematoxylin and eosin stained tissue. Advances in Intelligent Systems and Computing,
393–406. https://doi.org/10.1007/978-3-319- 91211-0_35
[6] Cascarano, Giacomo Donato, et al. “A Neural Network for Glomerulus Classification Based on
Histological Images of Kidney Biopsy.” BMC Medical Informatics and Decision Making, vol. 21,
no. S1, Apr. 2021, https://doi.org/10.1186/s12911-021-01650-
[7] Otálora, S., Atzori, M., Andrearczyk, V., Khan, A., & Müller, H. (2019). Staining invariant features
for improving generalization of deep convolutional neural networks in computational pathology.
Frontiers in Bioengineering and Biotechnology, 7. https://doi.org/10.3389/fbioe.2019.00198
[8] Pontalba, J. T., Gwynne-Timothy, T., David, E., Jakate, K., Androutsos, D., & Khademi, A.
(2019). Assessing the impact of color normalization in convolutional neural network-based nuclei
segmentation frameworks. Frontiers in Bioengineering and Biotechnology, 7.
https://doi.org/10.3389/fbioe.2019.00300
[9] Lakshmanan, B., et al. “Stain Removal through Color Normalization of Haematoxylin and Eosin
Images: A Review.” Journal of Physics: Conference Series, vol. 1362, no. 1, 1 Nov. 2019, p.
012108, https://doi.org/10.1088/1742-6596/1362/1/012108.
[10] Kang, Hongtao, et al. “Stainnet: A Fast and Robust Stain Normalization Network.” Frontiers in
Medicine, vol. 8, 5 Nov. 2021, https://doi.org/10.3389/fmed.2021.746307.
[11] Agraz, J. L., Grenko, C. M., Chen, A. A., Viaene, A. N., Nasrallah, M. D., Pati, S., Kurc, T., Saltz,
J., Feldman, M. D., Akbari, H., Sharma, P., Shinohara, R. T., & Bakas, S. (2022). Robust image
population based stain color normalization: How many reference slides are enough? IEEE Open
Journal of Engineering in Medicine and Biology, 3, 218–226.
https://doi.org/10.1109/ojemb.2023.3234443
[12] Ronneberger, O. (2017). Invited talk: U-net convolutional networks for biomedical image
segmentation. Informatik Aktuell, 3–3. https://doi.org/10.1007/978-3-662-54345-0_3
[13] Godwin, L. L., Ju, Y., Sood, N., Jain, Y., Quardokus, E. M., Bueckle, A., Longacre, T., Horning, A.,
Lin, Y., Esplin, E. D., Hickey, J. W., Snyder, M. P., Patterson, N. H., Spraggins, J. M., &
Börner, K. (2021). Robust and generalizable segmentation of human functional tissue units.
https://doi.org/10.1101/2021.11.09.467810
[14] Konovalenko, Ihor, et al. “Research of U-Net-Based CNN Architectures for Metal Surface Defect
Detection.” Machines, vol. 10, no. 5, 29 Apr. 2022, p. 327,
https://doi.org/10.3390/machines10050327.

105
[15] Botchkarev, A. (2019). A new typology design of performance metrics to measure errors in machine
learning regression algorithms. Interdisciplinary Journal of Information, Knowledge, and
Management, 14, 045–076. https://doi.org/10.28945/4184
[16] Li, X., Davis, R. C., Xu, Y., Wang, Z., Souma, N., Sotolongo, G., Bell, J., Ellis, M., Howell, D.,
Shen, X., Lafata, K. J., & Barisoni, L. (2021). Deep learning segmentation of glomeruli on
kidney donor frozen sections. Journal of Medical Imaging, 8(06).
https://doi.org/10.1117/1.jmi.8.6.067501
[17] Roy, S., kumar Jain, A., Lal, S., & Kini, J. (2018). A study about color normalization methods
for histopathology images. Micron, 114, 42–61. https://doi.org/10.1016/j.micron.2018.07.005
[18] Tellez D;Litjens G;Bándi P;Bulten W;Bokhorst JM;Ciompi F;van der Laak J; “Quantifying the
Effects of Data Augmentation and Stain Color Normalization in Convolutional Neural Networks for
Computational Pathology.” Medical Image Analysis, pubmed.ncbi.nlm.nih.gov/31466046/.
Accessed 31 July 2023.
[19] Wang, C.-M., Huang, Y.-H., & Huang, M.-L. (2006). An effective algorithm for image sequence
color transfer. Mathematical and Computer Modeling, 44(7-8), 608–627.
https://doi.org/10.1016/j.mcm.2006.01.029
[20] “3.1. Cross-Validation: Evaluating Estimator Performance.” Scikit, scikit-
learn.org/stable/modules/cross_validation.html. Accessed 5 Aug. 2023.
AUTHOR
Sai is a researcher associated with the university of Louisville's Biomedical Imaging
Lab. His concentrated intrest is exploring and applying new deep learning techniques to
biomedical data.

Evaluating the Impact of Color Normalization on Kidney Image Segmentation

More Related Content

Similar to Evaluating the Impact of Color Normalization on Kidney Image Segmentation (20)

Recently uploaded (20)

Evaluating the Impact of Color Normalization on Kidney Image Segmentation