0% found this document useful (0 votes)
44 views5 pages

10 UrinaryBladderCancer PDF

Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
44 views5 pages

10 UrinaryBladderCancer PDF

Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/361415957

Urinary Bladder Cancer Diagnostics Using Machine Learning Techniques

Conference Paper · June 2022

CITATIONS READS

0 49

3 authors:

Ivan Lorencin Josip Spanjol


University of Rijeka Klinički bolnički centar Rijeka
97 PUBLICATIONS 1,112 CITATIONS 111 PUBLICATIONS 995 CITATIONS

SEE PROFILE SEE PROFILE

Zlatan Car
University of Rijeka
238 PUBLICATIONS 1,501 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Development of an intelligent expert system for online diagnosis of bladder cancer View project

Cell Manufacturing View project

All content following this page was uploaded by Ivan Lorencin on 20 June 2022.

The user has requested enhancement of the downloaded file.


International Student Scientific Conference Ri-STEM 2022, Rijeka, Croatia, 8th-9th June 2022.

ISBN: 978-953-8246-26-5

Urinary Bladder Cancer Diagnostics Using Machine Learning Techniques


Ivan LORENCIN1, Josip ŠPANJOL2,3, Zlatan CAR4

1 Faculty of Engineering - University of Rijeka, Vukovarska 58, 51000 Rijeka, Croatia; [email protected]
2 Faculty of Medicine - University of Rijeka, Braće Branchetta 20/1, 51000, Rijeka, Croatia; [email protected]
2 Clinical Hospital Centre, Krešimirova ul. 42, 51000, Rijeka
4 Faculty of Engineering - University of Rijeka, Vukovarska 58, 51000 Rijeka, Croatia; [email protected]

Abstract: Urinary bladder cancer is one of the most common malignancies of the urinary system. Due to the high metastatic
potential and recurrence rate, artificial intelligence (ML) and machine learning (ML) methods are introduced to increase the
diagnostic performances. In this paper, a brief overview of the application of ML methods on two data sets is provided. In the
case of this research, cystoscopic images and images collected by using computerized tomography (CT) are used. From the
obtained results, it can be seen that utilization of ML methods enables higher results both in the case of classification and
semantic segmentation.

Keywords: artificial intelligence, clinical support systems, cystoscopy, computerized tomography, convolutional neural
network, data augmentation, edge detectors, meta-heuristic algorithms, transfer learning, urinary bladder cancer

1. Introduction

Urinary bladder cancer is one of the most common


malignancies of the urinary system. The root of cancer
lays in the mutation of the bladder mucosa cells. The main
characteristics of urinary bladder cancer are high
reoccurrence rate and high metastatic potential [1,2]. For
these reasons, it can be concluded that an accurate and fast
diagnostic procedure is of vital importance. With aim of
developing faster and more accurate diagnostic
procedures, artificial intelligence (AI) and machine
learning (ML) are introduced [3-5]. In this paper, an
overview of the AI utilization for urinary bladder cancer
diagnostics is provided. The presented diagnostic
procedures are based on two different diagnostic
procedures: a cystoscopy and computerized tomography. Figure 1: Example of each class from cystoscopy data set

2. Cystoscopy The presented data set is used to recognize the grade of


urinary bladder cancer by using AI-based classification
The first used data set is the data set collected during algorithms [6].
optical cystoscopy and examination of urinary bladder
mucosa with a confocal laser endomicroscope. The 3. Computerized Tomography
images of urinary bladder mucosa are divided into four
different classes: The other data set used in this research is the data set
• High-grade carcinoma, collected by using Computerized Tomography. The
• Low-grade carcinoma, images collected with CT are captured in three planes
• Carcinoma in-situ, and (frontal, horizontal, and sagittal) and are divided into 6
• Healthy mucosa classes (images without a bladder, healthy bladder,
unilateral bladder wall thickening, circular bladder wall
An example of each class is given in Figure 1. thickening, exophytic formation, and invasion outside the
contour of the bladder) [7]. An example for each plane
used in the research is given in Figure 2.

1
International Student Scientific Conference Ri-STEM 2022, Rijeka, Croatia, 8th-9th June 2022.

Figure 2: Example of each plane from CT data set


Figure 4: Schematic representation of U-net architecture
The aim of this data set is to develop a system for semantic
segmentation of urinary bladder cancer masses from CT
images. A brief process of semantic segmentation is given 4.3. Hybrid models
in Figure 3.
Alongside standard AI-based algorithms, several
approaches for hybrid models are developed as well. The
aim of hybrid model utilization is to increase not only
classification or semantic segmentation performances, but
also generalization performances. In this section, a brief
description of the hybrid models is presented.

4.3.1. Data Augmentation

With aim of performance-boosting, data augmentation is


used. For the case of cystoscopy data sets, both
geometrical and GAN-based augmentation procedures are
Figure 3: Semantic segmentation process used. For the case of CT data set, only geometrical
augmentation is used.
4. Methods overview
4.3.2. Edge detectors
In this section, a brief description of the used AI methods
is provided. With aim of performance-boosting, edge detector-based
hybrid models are also proposed. For this purpose, two
4.1 Classification different edge detection methodologies are utilized. The
first approach is to utilize gradient edge detectors. In this
For the classification of images contained in CT and research, three different gradient kernels are used [9,10]:
cystoscopy data sets, several standard CNN architectures
• Roberts edge detector,
are used [8]:
• Sobel edge detector, and
• AlexNet
• Prewitt edge detector.
• VGG-16
• ResNet50 Alongside gradient edge detector, Laplacian edge detector
• ResNet101 is used as well.
• ResNet152
• InceptionV3 4.3.3. Transfer Learning
• Inception-ResNet
One of the modern approaches in machine learning is
The aforementioned CNN architectures are also used as certainly transfer learning. In this research, transfer
pre-trained backbones for the implementation of transfer learning methodology is applied to both classification and
learning methodology in U-net for semantic segmentation. semantic segmentation. The transfer learning approach
includes the implementation of standard CNN
4.2. Semantic segmentation architectures pre-trained by using ImageNet data set. For
the case of classification problems, pre-trained
For the purposes of semantic segmentation, a standard U- architectures with added fully connected layers are used.
net architecture is used. A schematic representation of the On the other hand, in the case of U-net architecture, pre-
used U-net architecture is given in Figure 4. trained CNN architectures are used as contractive parts of
a U-net architecture [11-13].

4.3.4. Model selection using meta-heuristic approach

2
International Student Scientific Conference Ri-STEM 2022, Rijeka, Croatia, 8th-9th June 2022.

Alongside presented methods, meta-heuristic algorithms


are used for model selection of the proposed classification
and semantic segmentation algorithms. In this research,
two meta-heuristic algorithms are used:
• Genetic algorithm (GA) and
• Discrete Particle Swarm algorithm (D-PS).

Both algorithms are executed by using an adapted fitness


function that includes both classification and Figure 5: Classification of cystoscopic images using VGG-16
segmentation, as well as generalization performance. This
transfer function can be defined as:
5.2. Results achieved with CT data set
𝑭(𝝁)
̅̅̅̅̅̅̅𝟏 − 𝝈(𝚿(𝝁)),
= 𝚿(𝝁) (𝟏) If the results achieved on the CT data set are compared, it
can be seen that the highest semantic segmentation and
where: generalization performance are achieved if a general
semantic segmentation system is used. In this case, the
• 𝐹(𝜇) – fitness function,
highest results are achieved if the U-net model is selected
• Ψ(𝜇) – evaluation measure, and
by using the D-PS algorithm. Furthermore, it can be seen
• 𝜎 – standard deviation .
that similar results are achieved if one semantic
segmentation system is used for each diagnosis. In the
Mean evaluation measure and standard deviation are
case of horizontal and frontal planes, the highest results
measures derived from the k-fold cross-validation
are achieved if U-net models are selected by using D-PS.
procedure used during training [14].
On the other hand, in the case of the sagittal plane, the
highest performances are achieved if the transfer learning
5. Results
methodology is used. If one semantic segmentation
system is used for each diagnosis separately, significantly
In this section, a brief overview of the achieved results is
lower performances are achieved, as presented in Figure
provided.
6.
5.1. Results achieved with Cystoscopy data set

If the results achieved by using AlexNet-based hybrid


models on the cystoscopic data set are compared, it can be
noticed that the highest classification and generalization
performances are achieved if meta-heuristic algorithms
are used. Furthermore, it can be seen that the sightly lower
performances are achieved if the edge detector-based
hybrid model and GAN-based augmentation are used, as
presented in Figure 5.
Figure 6: Results achieved with each semantic segmentation methods

6. Conclusions

From the presented results it can be seen that the higher


performances are achieved in the case of both data sets if
a form of the hybrid model is used. Such a property is
particularly emphasized in the case of meta-heuristic
algorithms and data augmentation. In final, it can be stated
that there is a possibility for the implementation of ML
Figure 5: Classification of cystoscopic images using AlexNet methods in urinary bladder cancer diagnostics. The
utilization of ML methods enables higher results both in
The similar results can be observed if VGG-16 architecture is the case of classification and semantic segmentation.
used. In this case, the highest performances are achieved if D-
PS is used for the model selection, as presented in Figure 6. Acknowledgement

This research has been (partly) supported by the CEEPUS


network CIII-HR-0108, European Regional Development
Fund under the grant KK.01.1.1.01.0009
(DATACROSS), project CEKOM under the grant
KK.01.2.2.03.0004, Erasmus+ project WICT under the

3
International Student Scientific Conference Ri-STEM 2022, Rijeka, Croatia, 8th-9th June 2022.

grant 2021-1-HR01-KA220-HED-000031177, and Networks for Urinary Bladder Cancer Diagnosis Recognition
University of Rijeka scientific grant uniri-tehnic-18-275- From CT Imagery." 2021 IEEE 21st International Conference
1447. on Bioinformatics and Bioengineering (BIBE). IEEE, 2021.
[8] Alzubaidi, Laith, et al. "Review of deep learning: Concepts,
Bibliography CNN architectures, challenges, applications, future directions."
Journal of big Data 8.1 (2021): 1-74.
[1] Duty, B. D., and M. J. Conlin. "Principles of urologic [9] Lorencin, Ivan, et al. "Using multi-layer perceptron with
endoscopy." Campbell-Walsh Urology, 11th ed.; Elsevier: Laplacian edge detector for bladder cancer diagnosis." Artificial
Philadelphia, PA, USA (2016): 136-152. Intelligence in Medicine 102 (2020): 101746.
[2] Lerner, Seth P., et al. "Fluorescence and white light [10] Lorencin, Ivan, et al. "Edge detector-based hybrid artificial
cystoscopy for detection of carcinoma in situ of the urinary neural network models for urinary bladder cancer diagnosis."
bladder." Urologic Oncology: Seminars and Original Enabling AI Applications in Data Science. Springer, Cham,
Investigations. Vol. 30. No. 3. Elsevier, 2012. 2021. 225-245.
[3] Ikeda, Atsushi, et al. "Support system of cystoscopic [11] Baressi Šegota, Sandi, et al. "Semantic Segmentation of
diagnosis for bladder cancer based on artificial intelligence." Urinary Bladder Cancer Masses From CT Images: A Transfer
Journal of endourology 34.3 (2020): 352-358. Learning Approach." Biology 10.11 (2021): 1134.
[4] Eminaga, Okyaz, et al. "An Efficient Framework for Video [12] Wurm, Michael, et al. "Semantic segmentation of slums in
Documentation of Bladder Lesions for Cystoscopy: A Proof-of- satellite images using transfer learning on fully convolutional
Concept Study." (2022). neural networks." ISPRS journal of photogrammetry and
[5] Checcucci, Enrico, et al. "Applications of neural networks in remote sensing 150 (2019): 59-69.
urology: a systematic review." Current Opinion in Urology 30.6 [13] Sharma, Suvash, et al. "Semantic segmentation with
(2020): 788-807. transfer learning for off-road autonomous driving." Sensors
[6] Lorencin, Ivan, et al. "On urinary bladder cancer diagnosis: 19.11 (2019): 2577.
Utilization of deep convolutional generative adversarial [14] Lorencin, Ivan. An Intelligent System for Urinary Bladder
networks for data augmentation." Biology 10.3 (2021): 175. Cancer Diagnostics. Diss. University of Rijeka. Faculty of
[7] Lorencin, Ivan, et al. "Utilization of Convolutional Neural Engineering, 2022.

View publication stats

You might also like