Comparision Through Architectures of Semantic Segmentation in Breast

Ultrasound Images Across Differents Input Data Dimensions

ecio Elias Silva E. Silva

, Salom

ao Machado Mafalda

, Emili Silva Bezerra

Gustavo Moreira Oliveira de Castro

, Paulo Chavez dos Santos J

unior

and Ana Beatriz Alvarez

PAVIC Laboratory, University of Acre (UFAC), Rio Branco, Acre, Brazil

Keywords:

Breast Cancer, Semantic Segmentation, Neural Networks, Deep Learning.

Abstract:

Breast cancer is a problem that affects thousands of people every year, early diagnosis is important for the

treatment of this disease. Deep learning methods shows impressive results in identiﬁcation and segmentation

of breast cancer task. This paper evaluates the impact of input size images on three semantic segmentation

architectures applied to breast tumour ultrasound, in U-net, SegNet and DeepLabV3+. In order to (compre-

hensively) evaluate each architecture, 5-fold cross validation was carried out, thus reducing the impact of

variations in validation and training sets. In addition, the performance of the analyzed architectures was mea-

sured using the IoU and Dice metrics. The results showed that the DeepLabV3+ architecture performed better

than the others architectures in segmenting breast tumours, achieving an IoU of 0.70 and Dice of 0.73, with

the input dimension of the images being 128 × 128.

1 INTRODUCTION

Breast cancer is caused by the disorderly proliferation

of abnormal cells in the breast that form a tumor that

can invade other organs (Doi, 2007). Ultrasound has

been one of the most commonly used medical imag-

ing modalities for the assessment of breast masses.

It is also more accessible and cheaper than magnetic

resonance imaging (Athira et al., 2023).

The evaluation of ultrasound (US) breast images

requires extensive knowledge in this ﬁeld, since this

task involves identifying features of malignant and

benign breast masses on the images (Gharekhanloo

et al., 2018). To assist professionals in interpret-

ing these images, various computer-aided diagnosis

(CAD) systems have been developed (Doi, 2007).

The segmentation of nodules is of enormous impor-

tance in CAD systems, as its correct segmentation al-

lows a more accurate analysis of the morphological

features of breast tumors (Jahwar and Abdulazeez,

2022). However, automatic segmentation of US im-

https://orcid.org/0009-0000-1599-1167

https://orcid.org/0000-0002-2172-9402

https://orcid.org/0000-0003-4519-8332

https://orcid.org/0009-0002-4014-8803

https://orcid.org/0009-0004-5886-6541

https://orcid.org/0000-0003-3403-8261

ages poses a signiﬁcant challenge due to the low con-

trast of US images, the presence of speckle noise, and

the considerable variety in sizes and shapes of breast

nodules (Gharekhanloo et al., 2018).

Figure 1 illustrates the use of ultrasound images

to examine breast nodules. Figure 1.a shows an ultra-

sound image of a breast with a benign tumor, showing

a darker area (highlighted by the green square), with

well-deﬁned edges and shape; Figure 1.b shows an

image of a breast with a malignant tumor, with a dark

spot is highlighted (by the red rectangle), with unclear

shape and edges.

(a) Breast US with benign

tumor.

(b) Breast US with malig-

nant tumor.

Figure 1: Breast ultrasound images examples.

Accurate segmentation of breast masses in US im-

ages involves identiﬁcation and precise delineation of

354

Silva, C., Mafalda, S., Bezerra, E., Oliveira de Castro, G., Santos Júnior, P. and Alvarez, A.

Comparision Through Architectures of Semantic Segmentation in Breast Ultrasound Images Across Differents Input Data Dimensions.

DOI: 10.5220/0012322500003657

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

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 2, pages 354-361

ISBN: 978-989-758-688-0; ISSN: 2184-4305

the areas representing the masses, which allows more

detailed analysis of their characteristics (Yu et al.,

2021). One factor that can directly affect segmen-

tation quality and performance is the resolution of

US images, which varies depending on the equipment

used. The smaller the pixel size, the better ﬁne de-

tails can be captured, which can help in the accurate

segmentation of masses (Ploquin et al., 2015). On the

other hand, a smaller pixel size can result in a loss of

detail and affect the quality of the segmentation.

In addition, the low contrast in US images also

represents a challenge for automatic segmentation.

The presence of speckle noise in US images can

make it difﬁcult to correctly identify breast nodules

boundaries. The wide variation in size and shape

of breast masses also contributes to the complexity

of automatic segmentation in US images (Gokhale,

2009). This diversity requires robust and adaptable al-

gorithms that can handle different scenarios and guar-

antee reliable results (Ayana et al., 2022).

This paper investigates the inﬂuence of input im-

age size on the performance of semantic segmentation

neural networks applied to the most commonly used

state-of-the-art breast ultrasound images.

By investigating the inﬂuence of different input

image sizes on the performance of these architec-

tures, the research aims to provide valuable informa-

tion to optimize the segmentation process of breast

ultrasound images, which will ultimately contribute

to more accurate and reliable diagnoses of benign and

malignant breast cancer nodules.

2 DEEP NEURAL NETWORK

ARCHITECTURES

In recent years, several studies have been developed

for disease classiﬁcation, identiﬁcation, and segmen-

tation in medical images based on deep learning neu-

ral networks (Byra et al., 2020; Ayana et al., 2022).

Neural networks automatically extract features from

images related to some kind of anomaly, unlike clas-

sical digital image processing techniques (Sun et al.,

2022). The following architectures are most com-

monly used in the literature: U-net (Ronneberger

et al., 2015), DeepLabV3+ (Ayana et al., 2022) and

SegNet (Yurtkulu et al., 2019). This paper presents

an comparation through these three architectures.

2.1 U-Net

Proposed by (Ronneberger et al., 2015), U-Net is

a semantic segmentation neural network architecture

originally designed for medical image segmentation

tasks such as segmenting blood vessels and brain tu-

mors. U-Net has two main components that gives it

the “U” shape: the encoder path to map network in-

puts to feature map representations by convolutions

and max-pooling; and the decoder path to take that

feature representation as input using up-convolutions

and concatenations that are used for localizing object

boundaries. The architecture has 23 convolution lay-

ers, as illustrated in Figure 2.

Figure 2: An ilustration of U-Net archtecture. The gray ar-

rows indicate the skip connections that concatenate the en-

coder feature map with the decoder, which helps the back-

ward ﬂow of gradients for improved training. It consists of

a contracting path and an expansive path. The contracting

path follows the typical architecture of a convolutional net-

work.

The ability to handle limited amounts of data, ro-

bustness to variable image resolutions, and success-

ful application to complex segmentation tasks are fea-

tures of this architecture (Ronneberger et al., 2015).

2.2 SegNet

The SegNet architecture was proposed by (Badri-

narayanan et al., 2017), and is based on the VGG

family by the same author with the primary goal of

being used in scene understanding applications. With

a smaller total number of parameters than other archi-

tectures, it has better memory efﬁciency and speed.

Segnet has an encoder-decoder architecture followed

by pixel-level classiﬁcation. The advantage of SegNet

is in the way the decoder increases the sampling of its

Comparision Through Architectures of Semantic Segmentation in Breast Ultrasound Images Across Differents Input Data Dimensions

355

lower-resolution input feature maps. Speciﬁcally, the

decoder uses pooling indices computed in the max-

pooling phase of the corresponding encoder to per-

form non-linear upsampling. This eliminates the need

to learn how to increase resolution. In this way, high-

resolution details are preserved in the segmented im-

age (Badrinarayanan et al., 2017; Cui et al., 2023).

The Figure 3 illustrates the SegNet architecture.

Figure 3: An illustration of the SegNet architecture. There

are no fully connected layers and hence it is only convolu-

tional. A decoder upsamples its input using the transferred

pool indices from its encoder to produce a sparse feature

map(s). It then performs convolution with a trainable ﬁlter

bank to densify the feature map. The ﬁnal decoder output

feature maps are fed to a soft-max classiﬁer for pixel-wise

classiﬁcation.

2.2.1 DeepLabV3+

The DeepLab family was proposed by Google re-

searchers (Chen et al., 2017). The third version called

DeepLabV3 is an encoder-decoder network aimed at

performing semantic segmentation tasks. It is capa-

ble of capturing multi-contextual information in its

encoder by grouping features with multi-rate ﬁltering

and pooling operations that expand the receptive ﬁeld

of neurons, while the ﬁnal layers of the network can

capture sharper object boundaries, gradually recover-

ing spatial information. DeeplabV3+ adds a simple

but effective decoder module to DeepLabV3, added

to improve segmentation results (Liu et al., 2021).

In addition, DeepLabv3+ extends the Atrous Spa-

tial Pyramid Pooling module, which explores con-

volutional characteristics at various scales, applying

atrous convolutions at different rates (Yurtkulu et al.,

2019). The Figure 4 illustrates the DeepLabV3+ ar-

chitecture.

Figure 4: An ilustration of the DeeplabV3+ architecture.

The reason for using a Atrous Spatial Pyramid Pooling is

that it has been shown that as the sampling rate increases,

the number of valid ﬁlter weights (i.e., weights applied to

the valid feature region instead of padded zeros) becomes

smaller (Yurtkulu et al., 2019).

3 MATERIALS AND METHODS

3.1 BUS Dataset

The dataset used is called Breast Ultrasound Im-

ages (BUSI) and is made up of breast ultrasound im-

ages obtained from ultrasounds of 600 female patients

aged between 25 and 75. The BUSI dataset was pub-

lished in 2018 and has a collection of 780 images,

HEALTHINF 2024 - 17th International Conference on Health Informatics

356

with average dimensionality of 500x500 pixels. The

images were grouped by medical experts into three

categories: benign tumor, malignant tumor and nor-

mal images. Figure 5 shows examples of images from

BUSI dataset. In total, the dataset consists of 437 im-

ages of benign tumors and 210 images of malignant

tumors and 133 images with no tumors present (Al-

Dhabyani et al., 2020). In this article, only the images

with tumors present were used.

(a) An image of a benign

tumor.

(b) An image of a malig-

nant tumor.

tumor.

(d) Annotation of malig-

nant tumor.

Figure 5: Examples of breast ultrasound images from the

Dataset and their respective annotations.

3.2 Data Pre-Processing

For the neural networks semantic segmentation train-

ing, the annotations made by the experts were merged

into a single mask. Figure 6 shows the process for

merging 3 annotations of the tumors to form a single

output image.

In addition, in order to assess the dimensionality

impact of the input data on the segmentation perfor-

mance, ﬁve sets of data were created for training, dif-

fering only in the size of the images. The dimensions

established and analyzed in this article were: 32×32,

64×64, 128×128, 256×256 and 512× 512. Finally,

the data was divided into training, validation and test-

ing sets, following the proportion of 80%, 10% and

10% respectively.

Images of the malignant tumor class were aug-

mented to balance the dataset. The techniques used

were Horizontal Flip, Vertical Flip, Rotation 90º and

Transpose. 437 images were obtained for both the

malignant and benign classes, as shown in Figure 7.

Figure 6: Illustration of the annotation merging process.

Figure 7: Illustration graphically representing the total

number of images.

3.3 Deep Neural Network Training

The dimensionality of the data varies in different pro-

portions, from 32 to 512, and the architectures remain

with the same number of layers in all experiments.

For this reason, the regularization technique called

Early Stopping was used to avoid overfﬁting. This

method looks at the performance of the training data

and the validation set, this way when the model is ﬁt-

ting the training data and making errors in the vali-

dation data, i.e. not generalizing, the method deﬁnes

early stopping rules before the model starts to over-

ﬁt. This article used Early Stopping with a patience

of 20, i.e. the model was evaluated with the weights

20 epochs after overﬁtting.

The other hyperparameters used to train the archi-

tectures are shown in Table 1.

To illustrate the training process, Figure 8 shows

the loss and learning curve of the Segnet model for

the 64 × 64 dimension. A maximum of 100 epochs

was used for all training.

Comparision Through Architectures of Semantic Segmentation in Breast Ultrasound Images Across Differents Input Data Dimensions

357

Table 1: Hyperparameter table applied to training.

Hyperparameter Value

Learning rate 0.001

Batch-size 32

Epochs 100

Patience 20

Optimizer Adam

(a) Training and validation loss curve.

(b) Training and validation metrics.

Figure 8: Comparison of loss curve and metrics for training

and validation.

A K-Fold cross-validation (K = 5) was carried

out. This means that the data set was randomly di-

vided into ﬁve distinct parts, and the models were

trained and validated ﬁve times, alternating the train-

ing and validation parts in each iteration. This proce-

dure helps to evaluate the model’s performance more

comprehensively, reducing the impact of variations in

the training and validation sets and providing a reli-

able estimation of its ability to generalize (Bishop and

Nasrabadi, 2006).

For the results, the average of the performance

metrics obtained from cross-validation for each

method and input image size and class was calculated.

The differences in the results of the 5-Folds, for the

same input image size, showed an overall average of

approximately 0.06.

The low variance indicates that the models per-

form similarly in all parts of the dataset during the

cross-validation process, as proven by (Ostertagova

et al., 2014). Therefore, the results obtained in any of

the trained models can be considered representative

for all of them, providing a stable assessment of their

ability to generalize.

4 RESULTS AND DISCUSSION

This section presents the results obtained for the ex-

periments carried out for the three architectures eval-

uated: U-Net, SegNet and DeepLabV3+.

4.1 U-Net

Table 2 shows the results of the tests carried out with

the U-net architecture. The ﬁrst column shows the in-

put image sizes, the second column refers to the seg-

mented classes, while the last two columns show the

IoU and Dice. IoU and Dice for the benign tumor

class varies between 0.57 − 0.65 and 0.60 −0.67 with

a standard deviation between 0.32 − 0.42 and 0.34 −

0.40, respectively. For the malignant tumor class, IoU

and Dice varies between 0.31 − 0.56 and 0.35 − 0.60

with a standard deviation between 0.32 − 0.40 and

0.34 − 0.37, respectively.

Superior performance can be seen when the

model’s input image dimension is set to 64×64. With

IoU of 0.65 and Dice of 0.67 for the benign tumor

class and IoU of 0.56 and Dice of 0.60 for the ma-

lignant tumor class. Obtaining an average for both

classes of 0.60 for IoU and 0.64 for Dice.

Table 2: Segmentation performance scores (and standard

deviation) achieved by U-Net for different input image

sizes.

Size Class IoU Dice

32×32

Benign 0.65±0.38 0.67±0.36

Malignant 0.53±0.36 0.57±0.36

64×64

Benign 0.65±0.37 0.67±0.38

Malignant 0.56±0.35 0.60±0.35

128 × 128

Benign 0.63±0.40 0.65±0.40

Malignant 0.52±0.36 0.56±0.37

256 × 256

Benign 0.62±0.41 0.64±0.41

Malignant 0.39±0.40 0.44±0.37

512 × 512

Benign 0.57±0.42 0.60±0.42

Malignant 0.31±0.32 0.35±0.34

HEALTHINF 2024 - 17th International Conference on Health Informatics

358

Table 3: Segmentation performance scores (and standard

deviation) achieved by SegNet for different input image

sizes.

Size Class IoU Dice

32 × 32

Benign 0.59±0.42 0.60±0.45

Malignant 0.46±0.43 0.51±0.43

64 × 64

Benign 0.59±0.41 0.61±0.40

Malignant 0.57±0.38 0.61±0.38

128 × 128

Benign 0.61±0.39 0.64±0.39

Malignant 0.59±0.37 0.62±0.37

256×256

Benign 0.66±0.39 0.69±0.39

Malignant 0.59±0.38 0.63±0.38

512 × 512

Benign 0.65±0.38 0.68±0.39

Malignant 0.42±0.37 0.46±0.38

4.2 SegNet

Table 3 shows the results of the tests carried out with

the SegNet architecture. The ﬁrst column shows the

input image sizes, the second column refers to the

segmented classes, while the last two columns show

the IoU and Dice. IoU and Dice for the benign tumor

class varies between 0.59 − 0.66 and 0.60 −0.69 with

a standard deviation between 0.38 − 0.42 and 0.39 −

0.45, respectively. For the malignant tumor class, IoU

and Dice varies between 0.42 − 0.59 and 0.46 − 0.63

with a standard deviation between 0.37 − 0.43 and

0.37 − 0.43, respectively.

The Table 3 shows that the results obtained are

similar to those of U-Net. The standard deviation for

both benign and malignant classes was between 0.37

and 0.44 for IoU and 0.37 and 0.45 for Dice.

The architecture performed best in the 256 × 256

dimension. With IoU of 0.66 and Dice of 0.69 for the

benign and IoU of 0.59 and Dice of 0.63 for malignant

tumor class. The average IoU and Dice were 0.63 and

0.66, respectively.

4.3 DeepLabV3+

Table 4 shows the results of the tests carried out

with the DeepLabV3+ architecture. The ﬁrst col-

umn shows the input image sizes, the second col-

umn refers to the segmented classes, while the last

two columns show the IoU and Dice. IoU and Dice

for the benign tumor class varies between 0.68− 0.77

and 0.69 − 0.79 with a standard deviation between

0.33 − 0.40 for both metrics. For the malignant tu-

mor class, IoU and Dice varies between 0.51 − 0.63

and 0.53 − 0.66 with a standard deviation between

0.35 − 0.40 for both metrics.

The architecture performed best in the 128 × 128

dimension. With average IoU and Dice of 0.70 and

0.73, respectively.

Table 4: Segmentation performance scores (and standard

deviation) achieved by DeeplabV3+ for different input im-

age sizes.

Size Class IoU Dice

32 × 32

Benign 0.69±0.36 0.73±0.35

Malignant 0.54±0.36 0.61±0.36

64 × 64

Benign 0.68±0.38 0.71±0.38

Malignant 0.60±0.39 0.63±0.38

128×128

Benign 0.77±0.33 0.79±0.33

Malignant 0.63±0.35 0.66±0.35

256 × 256

Benign 0.73±0.37 0.74±0.37

Malignant 0.59±0.39 0.61±0.39

512 × 512

Benign 0.68±0.40 0.69±0.40

Malignant 0.51±0.40 0.53±0.40

4.4 Comparison Through Architectures

When comparing the three architectures imple-

mented, we can see a good performance in all archi-

tectures. U-net achieved the best performance with

an input size of 64 ×64, as shown in section 4.1. Seg-

Net achieved better performance with an input size of

256 × 256, as presented in section 4.2, with averages

of 0.60 for IoU and 0.64 for Dice. DeepLabV3+ per-

formed best with an input size of 128 ×128, as shown

in section 4.3, with averages of 0.63 and 0.66 for IoU

and Dice, respectively. The results obtained for the

DeepLabV3+ architecture were superior in all dimen-

sions tested for the other architectures, as shown in

Table 5, with an average of 0.70 and 0.73 for IoU and

Dice, respectively.

To illustrate the implementation of the three ana-

lyzed architectures, Figure 9 shows an example cre-

ated for U-net, SegNet, and DeepLabV3+. The ﬁrst

column shows the original image, the second column

shows the ground truth, and the third column the pre-

diction of the architecture. The ﬁgure shows a small

tumor with poorly deﬁned edges. In the ﬁgure, it

can be seen that DeepLabV3+ correctly segmented

the benign tumor (in orange), while SegNet misclas-

siﬁed the tumor, i.e., predicted it as a malignant tumor

(green) even though it was benign (orange). U-Net’s

prediction was inferior to the other two architectures

because it predicted two classes simultaneously. For

tumors with large proportions and well-deﬁned edges,

an example has been created and is shown in Figure

10. It can be seen that DeepLabV3+ and U-Net cor-

rectly segmented the benign tumor (in orange), while

SegNet prediction was inferior to the other two archi-

tectures because it predicted two classes simultane-

ously.

Comparision Through Architectures of Semantic Segmentation in Breast Ultrasound Images Across Differents Input Data Dimensions

359

Figure 9: Example of tumor segmentation with small tumor

with poorly deﬁned edges at different image sizes on the

analyzed architectures.

Figure 10: Example of tumor segmentation with large pro-

portions and well-deﬁned edges at different image sizes on

the analyzed architectures.

5 CONCLUSION

In this paper was evaluated the impact of the image

input size on semantic segmentation models applied

to breast tumor ultrasound, where the architectures

selected were U-net, SegNet and DeepLabV3+. The

architectures were evaluated using IoU and Dice. In

addition, to check the generalization capability of seg-

mentation architectures models was used independent

training-test experiments based on the k-fold cross-

validation method.

It was observed that the three architectures tested

performed differently in terms of image dimensions.

The U-net, Segnet and DeepLabV3+ architectures

performed best with input dimensions of 64 × 64,

256 × 256 and 128 × 128, respectively. These ﬁnd-

Table 5: Segmentation performance scores achieved by the

three architectures.

Size Architectures IoU Dice

32×32

U-net 0.59 0.62

SegNet 0.54 0.56

DeepLabV3+ 0.62 0.67

64×64

U-net 0.60 0.64

SegNet 0.58 0.61

DeepLabV3+ 0.64 0.67

128×128

U-net 0.57 0.61

SegNet 0.60 0.63

DeepLabV3+ 0.70 0.73

256×256

U-net 0.50 0.54

SegNet 0.63 0.66

DeepLabV3+ 0.66 0.68

512×512

U-net 0.44 0.47

SegNet 0.54 0.57

DeepLabV3+ 0.59 0.61

ings indicate that the sizing of input images is a crit-

ical factor in obtaining accurate and reliable results

in breast tumors segmentation. Each model showed

different preferences in relation to dimension, which

highlights the importance of a careful analysis of the

speciﬁc characteristics of each method when applying

them to different clinical and research contexts.

Furthermore, the signiﬁcant variation in IoU and

Dice results for different tumor classes emphasizes

the need to consider the peculiarities of each tumor

type when choosing the appropriate image dimension

for segmentation. This detailed understanding can op-

timize the accuracy of the segmentation process and,

consequently, improve the detection and follow-up of

breast tumors.

According to the results, DeepLabV3+, among all

the architectures analyzed, obtained the best perfor-

mance in semantic segmentation of breast tumors. As

expected, segmentation performance improves as net-

work depth increases. This behavior happens because

deeper networks extracted more subtle features that

distinguish normal and abnormal patterns in breast

tissue, while shallower networks extracted global im-

age features.

For approximately two decades, researchers de-

veloped CAD systems with high generalization capa-

bility that are useful in clinical practice. A compar-

ison of the different data input dimensions revealed

their potential to increase breast screening efﬁciency

and effectiveness. Hence, the experimental results en-

courage the use of DeepLabV3+ model within a CAD

system for the automated segmentation of breast tu-

mors.

Therefore, the appropriate choice of dimension is

a fundamental role in the effectiveness and accuracy

HEALTHINF 2024 - 17th International Conference on Health Informatics

360

of the predictions of semantic segmentation models.

This information is valuable for guiding future re-

search and clinical applications aimed at improving

the diagnosis and treatment of this important cancer

pathology.

ACKNOWLEDGEMENTS

The work presented in this paper was supported by

the Pesquisa Aplicada em Vis

ao e Intelig

encia Com-

putacional (PAVIC) project at Universidade Federal

do Acre, Brazil.

REFERENCES

Al-Dhabyani, W., Gomaa, M., Khaled, H., and Fahmy, A.

(2020). Dataset of breast ultrasound images. Data in

brief, 28:104863.

Athira, K., Dharmarajan, J. P., Vijaykumar, D., and Sub-

banna, N. (2023). Analysis of the various techniques

used for breast segmentation from mammograms. In

2023 International Conference on Distributed Com-

puting and Electrical Circuits and Electronics (ICD-

CECE), pages 1–7. IEEE.

Ayana, G., Dese, K., Raj, H., Krishnamoorthy, J., and Kwa,

T. (2022). De-speckling breast cancer ultrasound

images using a rotationally invariant block matching

based non-local means (ribm-nlm) method. Diagnos-

tics, 12(4):862.

Badrinarayanan, V., Kendall, A., and Cipolla, R. (2017).

Segnet: A deep convolutional encoder-decoder ar-

chitecture for image segmentation. IEEE transac-

tions on pattern analysis and machine intelligence,

39(12):2481–2495.

Bishop, C. M. and Nasrabadi, N. M. (2006). Pattern recog-

nition and machine learning, volume 4. Springer.

Byra, M., Jarosik, P., Szubert, A., Galperin, M., Ojeda-

Fournier, H., Olson, L., O’Boyle, M., Comstock, C.,

and Andre, M. (2020). Breast mass segmentation in

ultrasound with selective kernel u-net convolutional

neural network. Biomedical Signal Processing and

Control, 61:102027.

Chen, L.-C., Papandreou, G., Kokkinos, I., Murphy, K., and

Yuille, A. L. (2017). Deeplab: Semantic image seg-

mentation with deep convolutional nets, atrous convo-

lution, and fully connected crfs. IEEE transactions on

pattern analysis and machine intelligence, 40(4):834–

848.

Cui, W., Meng, D., Lu, K., Wu, Y., Pan, Z., Li, X., and Sun,

S. (2023). Automatic segmentation of ultrasound im-

ages using segnet and local nakagami distribution ﬁt-

ting model. Biomedical Signal Processing and Con-

trol, 81:104431.

Doi, K. (2007). Computer-aided diagnosis in medical imag-

ing: historical review, current status and future po-

tential. Computerized medical imaging and graphics,

31(4-5):198–211.

Gharekhanloo, F., Haseli, M. M., and Torabian, S. (2018).

Value of ultrasound in the detection of benign and ma-

lignant breast diseases: a diagnostic accuracy study.

Oman Medical Journal, 33(5):380.

Gokhale, S. (2009). Ultrasound characterization of breast

masses. Indian Journal of Radiology and Imaging,

19(03):242–247.

Jahwar, A. F. and Abdulazeez, A. M. (2022). Segmen-

tation and classiﬁcation for breast cancer ultrasound

images using deep learning techniques: a review. In

2022 IEEE 18th International Colloquium on Signal

Processing & Applications (CSPA), pages 225–230.

IEEE.

Liu, M., Fu, B., Xie, S., He, H., Lan, F., Li, Y., Lou,

P., and Fan, D. (2021). Comparison of multi-source

satellite images for classifying marsh vegetation using

deeplabv3 plus deep learning algorithm. Ecological

Indicators, 125:107562.

Ostertagova, E., Ostertag, O., and Kov

c, J. (2014).

Methodology and application of the kruskal-wallis

test. Applied mechanics and materials, 611:115–120.

Ploquin, M., Basarab, A., and Kouame, D. (2015). Res-

olution enhancement in medical ultrasound imaging.

Journal of Medical Imaging, 2(1):171.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-

net: Convolutional networks for biomedical image

segmentation. In Medical Image Computing and

Computer-Assisted Intervention–MICCAI 2015: 18th

International Conference, Munich, Germany, October

5-9, 2015, Proceedings, Part III 18, pages 234–241.

Springer.

Sun, Y., Yang, H., Zhou, J., and Wang, Y. (2022). Issmf:

Integrated semantic and spatial information of multi-

level features for automatic segmentation in prenatal

ultrasound images. Artiﬁcial Intelligence in Medicine,

125:102254.

Yu, K., Chen, S., and Chen, Y. (2021). Tumor segmenta-

tion in breast ultrasound image by means of res path

combined with dense connection neural network. Di-

agnostics, 11(9):1565.

Yurtkulu, S. C., S¸ahin, Y. H., and Unal, G. (2019). Se-

mantic segmentation with extended deeplabv3 archi-

tecture. In 2019 27th Signal Processing and Commu-

nications Applications Conference (SIU), pages 1–4.

IEEE.

Comparision Through Architectures of Semantic Segmentation in Breast Ultrasound Images Across Differents Input Data Dimensions

361