Content uploaded by Emanuela Genovese
Author content
All content in this area was uploaded by Emanuela Genovese on Oct 16, 2023
Content may be subject to copyright.
Bioengineering and Geomatics: Automatic Brain Image Segmentation
using Two-Stage Pipeline with SNN and Watershed Algorithm
VINCENZO BARRILE1*, EMANUELA GENOVESE1, ELENA BARRILE2
1Department of Civil, Energy, Environmental and Materials Engineering (DICEAM),
Mediterranea University of Reggio Calabria,
Via Graziella Feo di Vito- 89124, Reggio Cabria,
ITALY
2Vita-Salute San Raffaele University,
Via Olgettina, 58, 20132, Milan,
ITALY
*Corresponding Author
Abstract: - Digital image processing holds an increasingly essential role in the medical domain. This study
emphasizes the significance of researching and implementing methods aimed at the segmentation of critical
image regions and potential noise reduction, which is indispensable for medical professionals in disease
diagnosis. Consequently, the investigation of software solutions in this context can substantially enhance
diagnostic accuracy. In particular, neurology stands as a medical field wherein imaging plays a substantial
contributory role. In pursuit of an automated brain image segmentation approach, this paper centers its attention
on a two-step pipeline methodology to address the segmentation challenges inherent in medical imaging. The
proposed method incorporates the use of a Self-Normalizing Neural Network (SNN) for denoising and employs
the Watershed algorithm, typically employed in Geomatics imagery, for segmentation. Encouraging results are
obtained, with a segmentation performance, as measured by IoU, reaching a noteworthy value of 0.93 when
compared with alternative segmentation software.
Key-Words: - Neurology, biomedicine, neural networks, watershed technique, segmentation
Received: June 25, 2022. Revised: September 15, 2023. Accepted: October 5, 2023. Published: October 12, 2023.
1 Introduction
Segmentation in medical and geomatic images can
pose challenges due to image variations in contrast,
noise, and brightness, making it difficult to
distinguish regions accurately. Additionally, images
may contain artifacts like blurs, shadows, irregular
shapes, diversity between subjects, and other
complexities, which prolong and complicate the
segmentation process, [1], [2], [3]. As manual brain
segmentation is time-consuming, and segmenting
brain boundaries is complicated due to shadows and
noise, [4], [5], [6], machine learning and artificial
intelligence offer new segmentation techniques that
can recognize patterns in images and enhance image
quality while minimizing artifacts. There are several
segmentation techniques commonly used in medical
imaging:
1. Thresholding Technique: This method
involves comparing each pixel's intensity value with
a chosen threshold. If the pixel's intensity is lower
than the threshold, it's assigned to the background;
otherwise, it's assigned to the target region. The
result is a binary image, where pixels below the
threshold are set to 0, and those above it are set to 1.
However, this technique can only generate two
classes and does not consider spatial characteristics.
2. Region-Based Segmentation: Region-based
algorithms partition the image into similar regions
based on predefined criteria. One approach is
"region growing," where initial regions are defined,
and pixels are added to them if their intensity is
similar to the region's average value. Another
method is "region splitting and merging," which
assumes the image is initially a single region and
divides it into smaller regions if needed. Adjacent
regions that meet certain criteria are then merged.
3. Edge Detection: This method focuses on
recognizing contours in an image. It results in a
binary image where contours are assigned a value of
1, while the background is set to 0. Edge detection
relies on identifying changes in image intensity and
often uses derivative filters to estimate pixel
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.20
Vincenzo Barrile,
Emanuela Genovese, Elena Barrile
E-ISSN: 2224-2902
197
Volume 20, 2023
gradients. However, merely detecting edges may not
be sufficient to recognize significant regions in an
image, as many edges may be incomplete or
intersect.
4. Neural Networks: Neural network-based
segmentation is a departure from conventional
algorithms. It represents an image as a weighted
graph, where nodes correspond to one or more
pixels, and edge weights indicate similarity between
adjacent pixels. Various algorithms can effectively
partition these nodes to achieve segmentation, [7],
[8], [9].
Each of these segmentation techniques has its
strengths and limitations, and the choice of method
depends on the specific characteristics of the images
and the desired results in medical imaging
applications. In this context, this paper introduces a
new technique in this field: the watershed algorithm,
commonly used in geomatics for orographic
structure segmentation. However, the watershed
technique comes with certain limitations that
complicate its application in geomatic/medical
images such as the sensitivity to noise. For this
reason, an SNN-SELU neural network is used for
the denoising phase, guaranteeing the applicability
of the proposed methodology.
2 Materials and Methods
The proposed methodology consists of a two-stage
pipeline. The first stage involves the use of a
supervised learning neural network: SNN neural
network is modeled to operate the denoising phase
of the pipeline. The SNN must identify the noise
that if it is present in the brain structure makes it
difficult to apply the Watershed algorithm. The
second stage, instead, involves the application of the
watershed algorithm to the sharpened and processed
images, [10], [11].
2.1 First Stage: Denoising Process with Self-
Normalising Neural Network (SNN)
As known, a typical neural network consists of an
input layer and an output layer. The number of
neurons in the input layer depends on the specific
task and implementation choices, while the size of
the output layer varies based on the number of
desired output values for the model. The presence of
multiple neurons in the output layer can influence
the accuracy of the network's predictions. In
artificial neurons, inputs are combined with
corresponding weights to calculate a weighted sum
of the inputs. This weighted sum is then passed
through an activation function, which transforms it
into an output. Activation functions play a crucial
role in neural networks by mapping input data to
output values. This feature is essential for enabling
neural networks to learn intricate relationships and
patterns within datasets.
Self-normalizing neural networks (SNN), as
those used in the study, employ a specific
architecture where one of the layers is comprised of
neurons that use Scaled Exponential Linear Units
(SELU) as activation functions. SNNs are designed
to be robust against noise, and they possess the
unique characteristic of self-normalization, which
means they do not require extensive preprocessing
of input data to function effectively. One of the most
significant functions in this network is the SELU
activation function. SELU, short for Escalated
Exponential Linear Unit, plays a crucial role in
ensuring stable training and convergence during the
learning process. It offers the advantage of unit
variance in training errors and convergence toward
an average of zero. Notably, SELU is fast and does
not require complex initialization methods, making
it particularly effective when dealing with noisy
training datasets. SELU's properties also promote
self-normalization within the network, helping to
mitigate the issue of gradient disappearance during
training. The combination of Self-Normalizing
Neural Networks (SNN) and SELU activation is
highly advantageous when designing deep neural
networks. This combination ensures that gradients
remain stable throughout the training process,
enabling networks to learn intricate data
representations effectively. In comparison to other
activation functions like Rectified Linear Unit
(ReLU), SELU is often preferred, especially in
convolutional neural networks (CNNs), due to its
desirable properties and performance.
The SELU (Scaled Exponential Linear Unit)
activation function is defined as:
(1)
In this equation:
- `x` is the input to the SELU function.
- `alpha` is a constant with a value of approximately
1.6732.
- `scale` is a constant with a value of approximately
1.0507.
- `e` is the mathematical constant, approximately
equal to 2.71828.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.20
Vincenzo Barrile,
Emanuela Genovese, Elena Barrile
E-ISSN: 2224-2902
198
Volume 20, 2023
The SELU activation function is designed with
these specific values of `alpha` and `scale` to help
maintain unit variance in training errors and ensure
convergence towards an average of zero when
weights are correctly initialized. These constants
contribute to the effectiveness of SELU in deep
neural networks by addressing issues such as
vanishing gradients and promoting self-
normalization during training, [12], [13].
The neural network training phase is a crucial
point of the methodology. The initial step in
constructing the training dataset involves subjecting
a single "IMi" image to a custom pseudo-random
algorithm. This algorithm is crafted using the
OpenCV library in conjunction with functions from
the Computational Photography and denoising
package. The software is applied "n" times to the
same image, with variations occurring each time.
These variations are introduced randomly and
involve altering both the filter parameters provided
by the library and the sub-regions of the image
where the filter is most relevant. This process
generates a set of images labeled as "IMd11 ...
IMd1n." Each of these images serves as input for the
subsequent segmentation phase, which utilizes
OpenCV's Watershed algorithm. This segmentation
stage produces a set of images denoted as "IMs11 ...
IMs1n." These images contain the segmentation
results. To ensure the quality of the produced
images, a heuristic algorithm assesses whether they
exhibit any segmentation errors. If errors are
detected, the corresponding images are discarded,
while those without errors are included in the
training dataset. It's worth noting that there may be
instances of false positives, but the removal process
is error-free. The heuristic algorithm rejects images
when the generated polygons exhibit an unrealistic
number and area. For the elimination of false
positives, human intervention is necessary.
However, the operator's role is limited to a
verification process rather than manual
segmentation for network training. To address false
positives, a filtering operation was performed,
primarily focusing on eliminating glaring errors
without employing specific anatomical science-
based thresholds. The algorithm employed for
creating the training dataset follows a generate-and-
test approach. Despite the wide range of possible
permutations due to parameter variations, it remains
compatible with existing computational resources,
[14], [15].
In the final training dataset, each "IMi" image is
paired with a set of "IMd11 ... IMd1n" images that
have successfully passed the filtering and heuristic
assessment stages. The neural network's inference
must be directed towards a function capable of
either "flattening" or "highlighting" those pixels
whose noisy values could potentially disrupt the
effectiveness of the Watershed technique.
In this context:
- "Flattening" refers to the process of reducing
or equalizing the intensity values of noisy pixels,
making them less likely to interfere with the
Watershed segmentation.
- "Highlighting" involves enhancing or
emphasizing certain pixels, possibly those
representing important features or boundaries, to
improve the Watershed segmentation's accuracy.
The neural network's role is to process the input
images and produce an output that helps prepare the
data for successful segmentation using the
Watershed technique. Depending on the specific
characteristics of the input images and the noise
present, the network should learn to either mitigate
noise or enhance critical information, ensuring that
the subsequent segmentation process is more robust
and accurate.
2.2 Second Stage: Watershed Algorithm
The watershed-transformed segmentation technique
draws an analogy between a grayscale image and a
topographic relief map. In this analogy, each pixel's
gray level (f(x, y)) is interpreted as its elevation,
akin to geographical altitudes, [16], [17]. The
process is analogous to how water droplets behave
on such a topographic surface, following this
process:
Grayscale as Topographic Relief: Grayscale
images are treated as topographic maps, where pixel
intensities signify elevations. Lower-intensity
regions correspond to lower altitudes, while higher
intensities represent higher elevations.
Watershed Lines and Collection Basins: This
technique identifies "collection basins" within the
image, similar to geographical watersheds. These
basins correspond to local minima, essentially the
lowest points on the topographic surface. Watershed
lines are generated from these local minima, serving
as dividing lines that separate different regions or
objects within the image.
Contours and Object Representation: In the
context of image processing, these watershed lines
effectively outline the contours of objects within the
image. Each object is enclosed within its respective
collection basin, with watershed lines acting as
boundaries between them.
As known, the Watershed algorithm is based on
mathematical principles and processes. The key
steps involved the Gradient Calculation computing
the gradient of the image to highlight regions of
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.20
Vincenzo Barrile,
Emanuela Genovese, Elena Barrile
E-ISSN: 2224-2902
199
Volume 20, 2023
interest using gradient filters like Sobel or Scharr;
Marker Initialization defining markers or seeds to
indicate the regions you want to segment within the
image; Flood-Fill and Region Growing simulating
filling basins from these markers, and regions are
grown until they meet at watershed lines; Watershed
Lines based on the flooding process and used to
segment the image.
In this specific case, the formula for calculating
the gradient of an image using the Sobel filter is as
follows:
For the gradient in the x-direction:
(2)
For the gradient in the y-direction:
(3)
Where:
I(x, y) represents the intensity of the pixel in the
original image at position (x, y).
Gx is the gradient in the x-direction.
Gy is the gradient in the y-direction.
These formulas calculate the gradient of the
image in both the horizontal (x) and vertical (y)
directions using the Sobel filter. The resulting
gradient can be used to detect changes in image
intensity, which often correspond to edges or
regions of interest. Subsequent mathematical
relationships were applied to define the watershed
lines-finding algorithm.
This approach proves effective in detecting the
outlines of objects within the image. The Watershed
technique excels in detecting object boundaries.
This feature makes it valuable in applications that
require object recognition and analysis. For
example, robotics can be used to identify objects in
an environment and make decisions based on such
information.
However, the versatility of Watershed extends
beyond image processing alone. It is widely used in
geomatics for segmenting orographic structures
such as hills and mountains, demonstrating its
adaptability to a wide range of contexts.
Another advantage is computational speed.
When applied efficiently, the Watershed technique
can yield rapid results, making it suitable for
scenarios that demand fast segmentation. This
feature is particularly useful in emergencies or
applications requiring real-time processing.
It should be noted that, despite its numerous
advantages, the Watershed technique may present
some challenges. Among these, over-segmentation
and sensitivity to noise may require additional
attention. Consequently, it is often used in
combination with other techniques or algorithms to
enhance the quality of segmentation.
Below (Figure 1), is an example of the
application of the Watershed algorithm’s process
applied to bidimensional topographic images.
Fig. 1: Watershed process
2.3 Information Extraction from Images
Dataset (IXI Dataset)
The IXI dataset serves as a valuable resource for
brain image segmentation, [18]. It encompasses a
vast collection of magnetic resonance imaging
(MRI) scans from healthy subjects, spanning
different age groups, including both young and
elderly individuals. Researchers have developed
algorithms for automated brain segmentation based
on this extensive dataset. The IXI dataset comprises
nearly 600 MR images of healthy individuals, each
acquired using a protocol that includes T1, T2, and
PD-weighted images, MRA images, and diffusion-
weighted images with 15 different directions.
Figure 2 shows an example of an image acquired by
the IXI dataset.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.20
Vincenzo Barrile,
Emanuela Genovese, Elena Barrile
E-ISSN: 2224-2902
200
Volume 20, 2023
Fig. 2: IXI dataset: IXI263-HH-1684-T1_70
In the case study, 200 images from the IXI
dataset and some images provided by clinical
laboratories in the manner prescribed by law were
used. The initially generated training set contained
approximately 4,200 occurrences, but it was later
reduced to about 3,300 occurrences through manual
intervention to enhance data quality. The data
processing process was carried out independently on
several workstations, each equipped with eight 11th-
generation Intel Core i7 processors. Each
workstation processed one image at a time, and
processing was halted if it exceeded a 4-hour
computation time. Following the acquisition of the
training set, a neural network model based on the
Self-Normalizing Neural Network (SNN) with
SELU (Scaled Exponential Linear Unit) activation
was defined. This type of network is known for its
self-normalizing capability, which can contribute to
training stability and effectiveness, [19], [20], [21],
[22].
Subsequently, the supervised training phase of
the neural network was conducted using the
prepared dataset.
3 Results
After the execution of the training phase, the
network's performance within the pipeline
framework was assessed, essentially subjecting the
methodology to unit testing. To carry out this
evaluation, a set of 500 images that were not part of
the training set was employed and processed
through the pipeline. The main purpose of the
proposed method is not so much to accurately
identify brain structures, a task entrusted to medical
professionals, but rather to provide valuable support
during the diagnosis phases where time is limited.
This type of segmentation offers significant
advantages as it is cost-effective and allows for
quick results, which is crucial in emergencies.
However, it is important to emphasize that,
despite its speed and efficiency, this technique is
capable of segmenting some of the fundamental
brain structures, including the optic nerves, pituitary
gland, brainstem, and peduncle. In essence, it aims
to identify these key regions of the brain to facilitate
diagnosis and provide preliminary guidance to
medical professionals.
Figure 3 and Figure 4 show the whole pipeline
segmentation, denoising stage, and Watershed stage,
applied to cerebral images.
Fig. 3: Complete image processing pipeline,
including both the denoising and Watershed stages.
Fig. 4: Example of the whole processing pipeline
applied to another cerebral image.
To validate the results obtained, For the sake of a
comprehensive analysis, with a primary focus on
one of the advantages of the proposed method,
which is the reduced image segmentation time
compared to traditional methods, we conducted an
evaluation using similarity metrics. Specifically, we
measured the Intersection over Union (IoU) between
manual/ITK-SNAP (software used by experts in the
segmentation process) segmentations and the
proposed segmentation, [23], [24]. This assessment
involved examining different sections of the images
to quantify the degree of overlap between the two
segmentation approaches.
The results of the similarity analysis, which also
took into account false negatives and false positives,
revealed an average IoU of 0.93. This value
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.20
Vincenzo Barrile,
Emanuela Genovese, Elena Barrile
E-ISSN: 2224-2902
201
Volume 20, 2023
suggests that the images are generally comparable,
indicating strong agreement. However, it's worth
noting that there are variations in performance
across different regions, with some areas showing
excellent results while others may benefit from
further improvements.
4 Conclusion
The study explored the application of an approach
based on a neural network for image denoising,
followed by the utilization of the Watershed
algorithm for segmentation in the field of medical
images. This methodology was assessed to
comprehend its potential and limitations.
The results obtained from this research indicate
that the integration of a neural network for
denoising with the Watershed algorithm for
segmentation offers an intriguing perspective on the
management of medical images. However, it is
crucial to consider some critical aspects that
emerged from the data analysis. One of the main
challenges pertains to over-segmentation, a situation
where the Watershed algorithm divides image
regions into segments that are excessively small or
detailed. This can complicate result interpretation
and necessitate additional post-processing steps to
obtain coherent and clinically meaningful segments.
However, it should be emphasized that, despite
this challenge, the methodology presents numerous
advantages, including its effectiveness in detecting
object contours and its adaptability to various
medical applications. Regarding prospects, there are
several interesting research directions. The use of
larger and more representative training datasets
could enhance the denoising neural network's
capability.
In conclusion, this research lays the groundwork
for further developments in the field of medical
image segmentation through the combined use of
neural networks and the Watershed algorithm.
Despite the challenges, this methodology offers
significant opportunities to improve the quality and
efficiency of medical image analysis, with a
particular focus on image-assisted diagnosis.
References:
[1] Zhu, Y., Abdalla, A., Tang, Z., Cen, H.
(2022). Improving rice nitrogen stress
diagnosis by denoising strips in hyperspectral
images via deep learning, Biosystems
Engineering, Vol. 219, pp. 165-176.
[2] Taher, F., Mahmoud, A., Shalaby, A., & El-
Baz, A. (2018, December). A review on the
cerebrovascular segmentation methods. In
2018 IEEE International Symposium on
Signal Processing and Information
Technology (ISSPIT) (pp. 359-364). IEEE.
[3] Alirezaie, J., Jernigan, M. E., & Nahmias, C.
(1998). Automatic segmentation of cerebral
MR images using artificial neural networks.
IEEE Transactions on Nuclear Science, vol.
45(4), pp. 2174-2182.
[4] Zhu, J., Shi, H., Song, B., Tao, Y., Tan, S.,
Zhang, T. (2021). Nonlinear process
monitoring based on load weighted denoising
autoencoder, Measurement, Vol. 171, 108782.
[5] Tripathi, S., Sharma, N. (2021). Computer-
aided automatic approach for denoising of
magnetic resonance images. Computer
Methods in Biomechanics and Biomedical
Engineering: Imaging & Visualization, vol.
9:6, pp. 707-716.
[6] Angiulli, G., Barrile, V., & Cacciola, M.
(2005). SAR imagery classification using
multi-class support vector machines. Journal
of Electromagnetic Waves and Applications,
19(14), 1865-1872.
[7] Klambauer, G., Unterthiner, T., Mayr, A.,
Hochreiter, S. (2017). Self-Normalizing
Neural Networks. In Proceedings of the 31st
Conference on Neural Information Processing
Systems, Long Beach, CA, USA. Advances in
Neural Information Processing Systems 2017,
30.
[8] Barrile, V., Candela, G., & Fotia, A. (2019).
Point cloud segmentation using image
processing techniques for structural analysis.
The International Archives of the
Photogrammetry, Remote Sensing and Spatial
Information Sciences, vol. 42, pp. 187-193.
[9] Yushkevich P. A., Pashchinskiy A., Oguz I.,
Mohan S., Schmitt J. E., Stein J. M., Zukić D.,
Vicory J., McCormick M., Yushkevich N.,
Schwartz N., Gao Y., & Gerig G. (2019).
User-Guided Segmentation of Multi-modality
Medical Imaging Datasets with ITK-SNAP.
Neuroinform vol. 17, pp.83-102.
[10] Barrile, V., Cotroneo, F., Genovese, E., &
Bilotta, G. (2023). Using Snn Neural
Networks Trained with High Resolution Data
2 and Applied to Copernicus SENTINEL-2
Data. The International Archives of the
Photogrammetry, Remote Sensing and Spatial
Information Sciences, vol. 48, pp. 27-31.
[11] Barrile, V., Cotroneo, F., Genovese, E.,
Barrile, E., & Bilotta, G. (2023). An AI
Segmenter on Medical Imaging for Geomatics
Applications Consisting of a Two-State
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.20
Vincenzo Barrile,
Emanuela Genovese, Elena Barrile
E-ISSN: 2224-2902
202
Volume 20, 2023
Pipeline, Snns Network and Watershed
Algorithm. The International Archives of the
Photogrammetry, Remote Sensing and Spatial
Information Sciences, vol. 48, pp. 21-26.
[12] Wang, S., Yang, D.M., Rong, R., Zhan, X.,
Xiao, G. (2019). Pathology Image Analysis
Using Segmentation Deep Learning
Algorithms. Am J Pathol., vol. 189(9), pp.
1686-1698.
[13] Wang, R., Chen, S., Ji, C., Fan, J., Ye Li, Y.
(2022). Boundaryaware context neural
network for medical image segmentation.
Med. Image Anal., vol. 78, pp. 102395.
[14] Li, P., Jiang, X., Kambhamettu, C., Shatkay,
H. (2018). Compound image segmentation of
published biomedical figures. Bioinformatics
1;34(7):1192-1199.
[15] Li, H., Chen, C., Fang, S., Zhao, S. (2017).
Brain MR image segmentation using NAMS
in pseudo-color. Comput Assist Surg
(Abingdon)., 22(sup1):170-175.
[16] Khiyal, M. S. H., Khan, A., & Bibi, A. (2009).
Modified Watershed Algorithm for
Segmentation of 2D Images. Issues in
Informing Science & Information Technology,
6.
[17] Acharjya, P. P., Sinha, A., Sarkar, S., Dey, S.,
& Ghosh, S. (2013). A new approach of
watershed algorithm using distance transform
applied to image segmentation. International
Journal of Innovative Research in Computer
and Communication Engineering, 1(2), 185-
189.
[18] IXI Dataset – Brain Development. (n.d.).
https://brain-development.org/ixi-dataset/
[19] Barrile, V., Cacciola, M., D’Amico, S., Greco,
A., Morabito, F. C., & Parrillo, F. (2006).
Radial basis function neural networks to
foresee aftershocks in seismic sequences
related to large earthquakes. In Neural
Information Processing: 13th International
Conference, ICONIP 2006, Hong Kong,
China, October 3-6, 2006. Proceedings, Part II
13 (pp. 909-916). Springer Berlin Heidelberg.
[20] Fu, Y., Lei, Y., Wang, T., Curran W.J., Liu,
T., Yang, X. (2020). Deep learning in medical
image registration: a review. Phys Med Biol.
2020, 22;65(20):20TR01.
[21] Fu, Y., Lei, Y., Wang, T., Curran, W.J., Liu,
T., Yang, X. (2021). A review of deep
learning-based methods for medical image
multiorgan segmentation. Phys Med. Vol. 85,
pp.107-122.
[22] Chen, X., Wang, X., Zhang, K., Fung, K.M.,
Thai, T.C., Moore, K., Mannel, R.S., Liu, H.,
Zheng, B., Qiu, Y. (2022). Recent advances
and clinical applications of deep learning in
medical image analysis. Med Image Anal.,
vol. 79, pp.102444.
[23] Barrile, V., Bilotta, G., Fotia, A., & Bernardo,
E. (2020). Road extraction for emergencies
from satellite imagery. In Computational
Science and Its Applications–ICCSA 2020:
20th International Conference, Cagliari, Italy,
July 1–4, 2020, Proceedings, Part IV 20 (pp.
767-781). Springer International Publishing.
[24] Nowozin, S. (2014). Optimal decisions from
probabilistic models: the intersection-over-
union case. In Proceedings of the IEEE
conference on computer vision and pattern
recognition (pp. 548-555).
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed to the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The authors have no conflicts of interest to declare.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2023.20.20
Vincenzo Barrile,
Emanuela Genovese, Elena Barrile
E-ISSN: 2224-2902
203
Volume 20, 2023
