Applied Deep Learning Architectures for Breast Cancer Screening

Classification

Asma Zizaan

, Ali Idri

1,2

and Hasnae Zerouaoui

Modelling, Simulation and Data Analysis, Mohammed VI Polytechnic University, Benguerir, Morocco

Software Project Management Research Team, ENSIAS, Mohammed V University, Rabat, Morocco

Keywords: Computer-Aided Screening, Deep Convolutional Neural Networks, Transfer Learning, Breast Cancer

Screening, Mammogram Classification, Image Processing.

Abstract: Breast cancer (BC) became the most diagnosed cancer, making it one of the deadliest diseases. Mammography

is a modality used for early detection of breast cancer. The objective of the present paper is to evaluate and

compare deep learning techniques applied to mammogram images. The paper conducts an experimental

evaluation of eight deep Convolutional Neural Network (CNN) architectures for a binary classification of

breast screening mammograms, namely VGG16, VGG19, DenseNet201, Inception ResNet V2, Inception V3,

ResNet 50, MobileNet V2 and Xception. This evaluation was based on four performance metrics (accuracy,

precision, recall and f1-score), as well as Scott Knott statistical test and Borda count voting system. The data

was extracted from the CBIS-DDSM dataset with 4000 images. And results have shown that DenseNet201

was the most efficient model for the binary classification with an accuracy of 84.27%.

1 INTRODUCTION

One of the most diagnosed types of cancer among

women is breast cancer, with statistics showing that

one out of eight females will be diagnosed with breast

cancer in their lifetime (The American Cancer

Society medical and editorial content team, 2022). In

2020, breast cancer was reportedly diagnosed in 2.3

million women and has caused 685 000 deaths

globally (Breast Cancer, n.d.). Survival rates started

to appear promising in countries where early

detection programs were combined with multiple

treatment options to eradicate this invasive illness

(Coleman, 2017). Breast cancer can be detected early

through screening mammography, which is one of the

most common screening modalities of our time.

Medical image processing is defined as the use

and the investigation of image files of the human

body, typically collected from a Computed

Tomography (CT), Magnetic Resonance Imaging

(MRI) scanner, or another type of X-ray system using

computerized quantification and visualization tools.

The purposes of this analysis are to help search for or

diagnose disorders and guide medical procedures

such as surgery planning and treatments (McAuliffe

et al., 2001). Machine learning, specifically deep

learning, has sparked major interest in its application

to medical image processing due to its rapid progress.

Various previously published works applied deep

learning models to multiple medical fields such as

breast cancer and diabetic retinopathy (Lahmar &

Idri, 2022; Zerouaoui & Idri, 2022). For example,

(Zizaan & Idri, 2022): Shen et al. has constructed a

deep learning algorithm that may accurately identify

breast cancer instances on routine screening

mammograms, using an "end-to-end" training

strategy. This study showed promising results and is

trained to reach a high accuracy when applied to

similar mammogram datasets (Shen et al., 2019). In

the same manner, Agarwal et al. proposed a CNN

framework for automated mass detection in full-field

digital mammograms (FFDM) using VGG16,

Resnet50, and InceptionV3 (Agarwal et al., 2019).

The common downfall to the mentioned articles

is: (1) the lack of variety of the used DL architectures

for better comparison, and (2) the choice of

evaluation methods that is often limited to the

accuracy and area under curve (AUC).

Thus, the aim of this paper is to develop as well

as evaluate various DL techniques applied to the

binary classification of the CBIS-DDSM dataset of

BCS mammograms, namely VGG16, VGG19,

DenseNet201, Inception ResNet V2, Inception V3,

Zizaan, A., Idri, A. and Zerouaoui, H.

Applied Deep Learning Architectures for Breast Cancer Screening Classiﬁcation.

DOI: 10.5220/0011723700003393

In Proceedings of the 15th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2023) - Volume 3, pages 617-624

ISBN: 978-989-758-623-1; ISSN: 2184-433X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

617

ResNet 50, MobileNet V2 and Xception

architectures. The empirical evaluation is set over two

steps, four performance measures in the first place,

namely: accuracy, precision, recall, f1-score, and

secondly, statistical testing using the Scott-Knott test

and Borda count system in order to select the best

performing model out of the eight. Such evaluation

methods are used to compare, cluster ad rank DL

models (Idri et al., 2016; Ottoni et al., 2019;

Zerouaoui et al., 2021; Zerouaoui & Idri, 2022). The

results of this study are discussed over two research

questions (RQs):

(RQ1): What is the overall performance of DL

techniques in BCS binary classification?

(RQ2): Are there any DL techniques that

noticeably outperform the others?

Accordingly, the main contributions of this article

are: (1) the development of eight end-to-end CNN

architectures, (2) preparing the data by using the

Contrast Limited Adaptive histogram equalization

(CLAHE) (Pizer et al., 1987) technique and intensity

normalization, and finally (3) comparing the results

of each model and selecting the best performing

among them.

The following is a breakdown of the paper's

structure: Section 2 presents the related works. In

Section 3, lays out the background of the study. As

for section 4, it explains the data preparation process

as well as the configuration of the eight DL

techniques. Section 5 presents and discusses the

empirical results. And finally, section 6 outlines the

conclusions and future works.

2 RELATED WORKS

This section summarizes the findings of previous

studies investigating deep learning techniques for

BCS classification. Zizaan et al. published a

systematic mapping study (SMS) of the use of

machine learning in BCS (Zizaan & Idri, 2022). A

total of 66 papers published between 2011 and 2021

were selected for analysis, the main findings of the

SMS were:

(1) The most common publication venue is

journals (58.88%), followed by conferences (6.9%),

and books (2.3%). These publications showed a peak

in 2019.

(2) The most frequent paper type is evaluation

research (92.4%) including 31.8% of solution

proposal papers. And the least frequent articles are

reviews (7.5%).

(3) The most used BCS modality is

mammography (61%), followed digital breast

tomosynthesis, MRI, and ABUS imaging (10%, 7%,

and 7% respectively).

Out of the sixty-six selected papers, nine articles

used the DDSM dataset. Table 1. presents the five

most recent studies and their findings. The most used

DL techniques were VGG16 and Resnet50, while the

Table 1:

Overview of five related studies.

Authors DL techniques

Performance

metrics

Results

Shen et al. VGG16, Resnet50,

Hybrid networks

Accuracy, AUC On several mammography platforms, automatic

deep learning methods can be easily trained to

achieve excellent accuracy.(Shen et al., 2019).

Aboutalib et

al.

AlexNe

AUC The DDSM dataset had the overall best

performance. This could be as a result of the

higher dataset size or the features of the dataset

itself. (Aboutalib et al., 2018).

Chougrad et

al.

ImageNet, VGG16,

RESNET50, and

inceptionv3

AUC, Accuracy Achieve 97.35% accuracy as well as 0.98 AUC

on the DDSM dataset (Chougrad et al., 2018)

Saranyaraj et

al.

Deep Convolutional

neural network

(DCNN), ResNet,

GoogleNet and VGG

Accuracy,

specificity, AUC,

recall, precision,

and F1-score

Increase the mean classification accuracy to

97.46% and the overall classification accuracy

to 96.23% (Saranyaraj et al., 2020)

Agarwal et al. VGG16, ResNet50, and

InceptionV3.

Accuracy Develop an automated framewor

that has

obtained the best results based on TPR and FPI

(Agarwal et al., 2019)

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

618

most common performance metrics are accuracy and

area under curve (AUC). Some of the noticeable

limitations were the low number of applied DL

techniques, averaging at three architectures, as well

as the performance metrics and the lack of statistical

tests.

3 EXPERIMENT

CONFIGURATION

This section presents the data pre-processing tasks,

the parameter tuning of the DL models, as well as the

empirical design.

3.1 Data Preparation

This dataset consists of images from CBIS-DDSM

dataset (Eric A. Scuccimarra, 2018). Digital Imaging

and Communications in Medicine format (DICOM)

pictures totaling 10,239 were collected from 1,566

patients across 6,775 trials to build this dataset (Lee

et al., 2017). Some Pre-Processing tasks have already

been done such as extracting the region of interest

(ROIs) and resizing to 299x299. The primary dataset

contains 55890 training mammograms, with 7825

(14%) positive cases and 48064 (86%) negative

cases.

Considering that more than half of the images in

the CBIS-DDSM dataset were labelled negative,

which is considered as a limitation, resampling the

images from the DDSM dataset was made by using

data augmentation. It is worth mentioning that the

CBIS-DDSM dataset is a subset of the DDSM

dataset, meaning that the images that were sampled

from DDSM are of the same nature as the images

existing in the CBIS-DDSM dataset.

Further image processing was made using the

Contrast Limited Adaptive histogram equalization

(CLAHE) technique to enhance the images’ contrast

as well as intensity normalization. In Figure 1, the

data preparation process is detailed: First, acquiring

the data from the DDSM and CBIS-DDSM datasets

with the same number of samples in each class. Then,

enhancing the contrast of the images using the

CLAHE technique. Lastly, applying the min-max

normalization as shown in Equation 1 to the input

photos will normalize them to the conventional

distribution. Results of the image enhancing

technique are shown in Figure 2.

𝑥





 





 





(1)

3.2 Model Configurations

As the present study is a binary classification of BCS

mammogram data extracted from the publicly

available dataset CBIS-DDSM, eight different end-

to-end CNN architectures were implemented using

several parameter tuning experiments. To train the

models after the data preparation phase was finished,

the transfer learning technique served as a practical

tool to import models pre-trained in the ImageNet

dataset (Fei-Fei et al., 2010).

To summarize the parameter tuning, the batch

size was set to 32 and the number of epochs to 200.

As for the optimization, the output layer optimization

function was SoftMax, and the optimizer Adam

(adaptive moment estimation) was chosen (Kingma

& Ba, 2015) with an initial learning set to 0. 00001.

Moreover, the loss function was set as the cross

entropy, and all the layers were frozen for the transfer

learning.

Figure 1: Data preparation process.

Figure 2: Example of CLAHE image transformation.

3.3 Data Splitting and Performance

Metrics

In this study, we used K-fold cross validation with k

equal to 5 to apply and evaluate the DL models. We

also provided the performance metrics' average for

each DL method. Four metrics—accuracy, precision,

recall, and f1-score—were used to evaluate the

performance of the eight DL classifiers. These

metrics are provided by Equations 2 through 5.

Applied Deep Learning Architectures for Breast Cancer Screening Classiﬁcation

619

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 





(2)

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 





(3)

𝑅𝑒𝑐𝑎𝑙𝑙 





(4)

𝐹1 𝑠𝑐𝑜𝑟𝑒2 





(5)

where:

• TP: true positives.

• FP: false positives.

• TN: true negatives.

• FN: false negatives.

In addition to the four performance metrics, two

statistical tests were used:

Scott Knott test is a type of exploratory clustering

techniques that is commonly employed in the context

of analysis of variance (ANOVA). Scott and Knott

proposed it in 1974 as a way to distinguish

overlapping groups using numerous comparisons of

treatment means (Jelihovschi et al., 2014).

Borda count is a voting method for single winner

election methods. Candidates are given points based

on their ranking in this method: 1 point for last choice,

2 points for second-to-last choice, and so on. The

entire point totals for all ballots are added together,

and the candidate with the highest point total wins

(García-Lapresta & Martínez-Panero, 2002).

3.4 Experimental Design

The methodology used to conduct the empirical

evaluations consists of three steps as shown in Fig. 3.

(1) Assess four performance metrics of each variant

of the deep learning architectures (VGG16,

VGG19, DenseNet201, Inception ResNet V2,

Inception V3, ResNet 50, MobileNet V2 and

Xception) as well as a CNN baseline.

(2) Select the DL architectures that outperformed the

CNN baseline by comparing the accuracy results.

(3) Use the Skott Knott test to build clusters of the

selected DL models and select the best SK cluster.

(4) To choose the ideal DL architecture, rank the top

SK cluster using the Borda count voting technique

based on the four performance criteria (accuracy,

precision, recall, and F1-score).

It is worth noting that comparable approaches were

utilized in previous works. (Azzeh et al., 2017 ; Idri

et al., 2018 ; Idri & Abnane, 2017 ; Worsley, 1977 ;

Zerouaoui et al., 2021)

4 RESULTS AND DISCUSSION

This section presents and discusses the evaluations'

findings for the eight DL approaches used on the

CBIS-DDSM dataset. The performances of the DL

techniques were evaluated using four numerical

performance metrics: accuracy, recall, precision, and

F1-score, as well as two statistical tests: Borda count,

and Scott-Knott. First, the performance of each DL

approach is examined, and the ones that outperform

the CNN baseline in accuracy are chosen. (RQ1).

Then, we utilize the Borda count voting

mechanism to rank the DL approaches belonging to

the best SK cluster after using the SK statistical test

based on accuracy to cluster the chosen DL

techniques (i.e., accuracy greater than the CNN

baseline model) (RQ2).

Figure 3:

Experimental design.

4.1 Assessment of the DL Models’

Performance

Eight different CNN architectures which were

previously pre-trained on ImageNet (VGG19,

VGG16, ResNet50, Inception V3, DenseNet201,

MobileNet V2, Xception, and Inception ResNet V2)

are applied over the DDSM dataset in Python using

CBIS-DDSM Dataset

VGG16, VGG19, DenseNet201, Inception

ResNet V2, Inception V3, ResNet 50, MobileNet

V2 and Xception

Step 1: Assess the performance of each DL model as well

as a CNN baseline by using four performance metrics

(Accuracy, specificity, recall, and F1-score)

Step 2: Select the DL architectures that outperformed the

CNN baseline by comparing the accuracy results.

Step 3: Use the Skott Knott test to build clusters of the

selected DL models and select the best SK cluster.

Step 4: Rank the best Scott Knott cluster using the Borda

count voting system using accuracy, precision, recall, F1-

score to select the best performing DL architecture.

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

620

Tensorflow (Abadi et al., 2016), Keras, SciKit-Learn

(Pedregosa et al., 2011), Pandas(Reback et al., 2022),

Matplotlib (Hunter, 2007), NumPy, and Seaborn

(Waskom, 2021) frameworks, on a tensor processing

unit (TPU) provided by Google in collab notebook.

Figure 4 and Table 2 show the accuracy values of

the DL models as well as the CNN baseline. These

results are summarized in Table 2 and reveal that:

a) DenseNet201 unlocks better performance than

the other CNN architectures with 84.27%

accuracy and 84.27% F1 score.

b) ResNet50 was the weakest model showing a

noticeably lower accuracy (76.35%) and F1-

score (76.46%).

c) MobileNetV2 outperforms Inception V3 by a

2.67% higher accuracy.

Table 2: Performance metrics.

CNN

Architecture

Accuracy Precision Recall

Score

CNN baseline 82.07% 80.98% 84.25% 82.50%

VGG19 81.00% 80.76% 81.85% 81.14%

VGG16 82.82% 81.48% 85.19% 83.26%

InceptionV3 80.75% 80.77% 81.10% 80.87%

DenseNet201 84.27% 84.55% 84.0% 84.27%

MobileNetV2 83.42% 83.47% 83.55% 83.48%

Xception 82.57% 82.64% 82.60% 82.61%

InceptionResne

tV2

83.12% 83.34% 82.65% 83.14%

ResNet50 76.35% 76.27% 76.92% 76.46%

Figure 4: Validation and training accuracy values of each

DL model.

Hereafter, the models to be discussed are the ones

that have scored a better performance than the CNN

baseline, namely: VGG16, DenseNet201,

MobileNetV2, Xception, and InceptionResNetV2.

4.2 Selection of the Best Performing

DL Model

Based on the results obtained, and to guarantee a

more precise selection of the best group of

architectures, further testing was required. In

particular, the hierarchical clustering of Scott Knott

was applied to the DL architectures which scored a

higher accuracy than the CNN baseline.

The outcome of the SK test shows that the

selected DL models all belong to a single cluster

(figure 5). As a result, these four DL approaches have

similar prediction skills in terms of accuracy values.

This cluster contains DenseNet201, MobileNet V2,

InceptionResNet V2, Xception, and VGG16.

Finally, selecting the best model from the Scott

Knott (table 3) cluster has been made by the Borda

count voting system taking into consideration the four

metrics (accuracy, recall, precision, and f1-score).

Figure 5: Results of Scott Knott test of the DL techniques.

Table 3: Borda count rankings.

DL Model Rank

DenseNet201 1

MobileNet V2 2

VGG16 3

Inception ResNet V2 4

Xception 5

5 THREATS TO VALIDITY

This section discusses the threats to the validity of this

paper from both external and internal perspectives.

5.1 Internal Validity

In order to strengthen the robustness of the mean

accuracy of the eight DL designs, this work applied

Applied Deep Learning Architectures for Breast Cancer Screening Classiﬁcation

621

the K-fold cross validation approach. When it comes

to optimization, the Adam (adaptive moment

estimation) optimizer's superior performance and

quick learning rate convergence make it preferable to

the traditional stochastic gradient descent method

(Zhang, 2019).

5.2 External Validity

The CBIS-DDSM dataset, which provides screening

mammography images, was utilized in this

investigation; however, we are unable to extend the

findings to other datasets that have the same image

type and attributes. In light of this, it would be

beneficial to repeat this experiment using more DL

methods, such as the UNET model or various CNN

model types, using additional publicly or privately

available datasets in order to corroborate or refute the

findings of the study.

5.3 Construct Validity

To assess the dependability of the classifier, this study

concentrated on the accuracy and other performance

measures (precision, recall and F1-Score). The fact

that these metrics are often used to assess

categorization performance was the primary factor in

the selection of these performance criteria. In order to

avoid favoring one performance criterion over

another, the results were also obtained using the SK

test and Borda count voting technique with equal

weights utilizing the four performance criteria.

6 CONCLUSIONS AND FUTURE

WORK

This paper presented and discussed the process and

the results of an empirical evaluation study of eight

end-to-end CNN architectures, notably VGG19,

VGG16, ResNet50, Inception V3, DenseNet201,

MobileNet V2, Xception, and Inception ResNet V2

for the binary classification of BCS mammogram

images retrieved from the datasets CBIS-DDSM and

DDSM. The empirical testing was based on four

performance metrics (accuracy, precision, recall, and

f1-score), as well as a statistical testing which

consisted of Scott-Knott test and Borda count voting

system. This study's findings can be summarized in

two main points:

(RQ1) What is the overall performance of DL

techniques in BCS binary classification?

The accuracy percentages of the eight DL

techniques were all satisfactory and the majority

exceeded 80%. The highest accuracies were scored

by DenseNet201, MobileNet V2, and Inception

Resnet V2, respectively. While ResNet50 was,

compared to the other models, underperforming.

(RQ2): Are there any DL techniques that

noticeably outperformed the others?

SK test resulted in only one cluster, so further

testing was done to ensure that the best model is

chosen. So, by ranking first in the Borda count voting

system, DenseNet201 is the best performing DL

architecture out of the eight trained DL models. Thus,

it is a highly recommended model to serve as base for

a DL computer assistance program to aid in the

process of BCS.

This work came as a part of a project that puts

efforts in building a comprehensive tool of computer

aided diagnosis (CAD) to help improve the process of

breast cancer screening in terms of radiologists’

assessment as well as reducing unnecessary steps of

the process to eventually result in an early diagnosis

of the disease.

As future work, we aim to experiment with

different ensemble techniques, in particular the

techniques that use bagging or boosting, and use a

variety of pre-processing steps aimed for feature

selection on breast cancer screening data. Moreover,

we plan to apply and evaluating different DL

architectures over tabular breast cancer screening

data as well.

ACKNOWLEDGMENTS

This work was conducted under the research project

“Machine Learning based Breast Cancer Diagnosis

and Treatment,” 2020-2023. The authors would like

to thank the Moroccan Ministry of Higher Education

and Scientific Research, Digital Development

Agency (ADD), CNRST, and UM6P for their

support.

REFERENCES

Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z.,

Citro, C., Corrado, G. S., Davis, A., Dean, J., Devin,

M., Ghemawat, S., Goodfellow, I., Harp, A., Irving, G.,

Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kudlur,

M., … Research, G. (2016). TensorFlow: Large-Scale

Machine Learning on Heterogeneous Distributed

Systems. https://arxiv.org/abs/1603.04467v2

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

622

Aboutalib, S. S., Mohamed, A. A., Berg, W. A., Zuley, M.

L., Sumkin, J. H., & Wu, S. (2018). Deep learning to

distinguish recalled but benign mammography images

in breast cancer screening. Clinical Cancer Research,

24(23), 5902–5909. https://doi.org/10.1158/1078-

0432. CCR-18-1115

Agarwal, R., Diaz, O., Lladó, X., Yap, M. H., & Martí, R.

(2019). Automatic mass detection in mammograms

using deep convolutional neural networks. Journal of

Medical Imaging, 6(3), 031409. https://doi.org/10

.1117/1.JMI.6.3.031409

Azzeh, M., Nassif, A. B., & Minku, L. L. (2017). An

empirical evaluation of ensemble adjustment methods

for analogy-based effort estimation. Journal of Systems

and Software, 103, 36–52. https://doi.org/10.

1016/j.jss.2015.01.028

Chougrad, H., Zouaki, H., & Alheyane, O. (2018). Deep

Convolutional Neural Networks for breast cancer

screening. Computer Methods and Programs in

Biomedicine, 157, 19–30. https://doi.org/10.1016/J.

CMPB.2018.01.011

Coleman, C. (2017). Early Detection and Screening for

Breast Cancer. Seminars in Oncology Nursing, 33(2),

141–155. https://doi.org/10.1016/J.SONCN.2017.02.

009

Eric A. Scuccimarra. (2018). DDSM Mammography.

Kaggle.

Fei-Fei, L., Deng, J., & Li, K. (2010). ImageNet:

Constructing a large-scale image database. Journal of

Vision, 9(8), 1037–1037. https://doi.org/10.1167/9

.8.1037

García-Lapresta, J. L., & Martínez-Panero, M. (2002).

Borda Count Versus Approval Voting: A Fuzzy

Approach. Public Choice 2002 112:1, 112(1), 167–184.

https://doi.org/10.1023/A:1015609200117

Hunter, J. D. (2007). Matplotlib: A 2D graphics

environment. Computing in Science and Engineering,

9(3), 90–95. https://doi.org/10.1109/MCSE.2007.55

Idri, A., & Abnane, I. (2017). Fuzzy Analogy Based Effort

Estimation: An Empirical Comparative Study. IEEE

CIT 2017 - 17th IEEE International Conference on

Computer and Information Technology, 114–121.

https://doi.org/10.1109/CIT.2017.29

Idri, A., Abnane, I., & Abran, A. (2018). Evaluating Pred(p)

and standardized accuracy criteria in software

development effort estimation. Journal of Software:

Evolution and Process, 30(4), e1925.

https://doi.org/10.1002/SMR.1925

Idri, A., Hosni, M., & Abran, A. (2016). Improved

estimation of software development effort using

Classical and Fuzzy Analogy ensembles. Applied Soft

Computing, 49, 990–1019.

https://doi.org/10.1016/J.ASOC.2016.08.012

Jelihovschi, E., Faria, J. C., & Allaman, I. B. (2014).

ScottKnott: A Package for Performing the Scott-Knott

Clustering Algorithm in R. TEMA (São Carlos), 15(1),

003. https://doi.org/10.5540/TEMA.2014.015.01.0003

Kingma, D. P., & Ba, J. L. (2015). Adam: A method for

stochastic optimization. 3rd International Conference

on Learning Representations, ICLR 2015 - Conference

Track Proceedings, 1–15.

Lahmar, C., & Idri, A. (2022). On the value of deep learning

for diagnosing diabetic retinopathy. Health and

Technology, 12(1), 89–105. https://doi.org/10.10

07/S12553-021-00606-X/FIGURES/11

Lee, R. S., Gimenez, F., Hoogi, A., Miyake, K. K.,

Gorovoy, M., & Rubin, D. L. (2017). A curated

mammography data set for use in computer-aided

detection and diagnosis research. Scientific Data 2017

4:1, 4(1), 1–9. https://doi.org/10.1038/sdata.2017.177

McAuliffe, M. J., Lalonde, F. M., McGarry, D., Gandler,

W., Csaky, K., & Trus, B. L. (2001). Medical image

processing, analysis & visualization in clinical

research. Proceedings of the IEEE Symposium on

Computer-Based Medical Systems, 381–388.

https://doi.org/10.1109/CBMS.2001.941749

Ottoni, A. L. C., Nepomuceno, E. G., de Oliveira, M. S., &

de Oliveira, D. C. R. (2019). Tuning of reinforcement

learning parameters applied to SOP using the Scott–

Knott method. Soft Computing, 24(6), 4441–4453.

https://doi.org/10.1007/S00500-019-04206-W

Pedregosa, F., Michel, V., Grisel OLIVIERGRISEL, O.,

Blondel, M., Prettenhofer, P., Weiss, R., Vanderplas, J.,

Cournapeau, D., Pedregosa, F., Varoquaux, G.,

Gramfort, A., Thirion, B., Grisel, O., Dubourg, V.,

Passos, A., Brucher, M., Perrot andÉdouardand, M.,

Duchesnay, A., & Duchesnay EDOUARD

DUCHESNAY, Fré. (2011). Scikit-learn: Machine

Learning in Python. Journal of Machine Learning

Research, 12, 2825–2830. https://doi.org/10.5555/1

953048

Pizer, S. M., Amburn, E. P., Austin, J. D., Cromartie, R.,

Geselowitz, A., Greer, T., ter Haar Romeny, B.,

Zimmerman, J. B., & Zuiderveld, K. (1987). Adaptive

histogram equalization and its variations. Computer

Vision, Graphics, and Image Processing, 39(3), 355–

368. https://doi.org/10.1016/S0734-189X(87)80186-X

Reback, J., jbrockmendel, McKinney, W., Bossche, J. van

den, Augspurger, T., Cloud, P., Hawkins, S., Roeschke,

M., gfyoung, Sinhrks, Klein, A., Hoefler, P., Petersen,

T., Tratner, J., She, C., Ayd, W., Naveh, S., Darbyshire,

J., Garcia, M., … Seabold, S. (2022). pandas-

dev/pandas: Pandas 1.4.1. https://doi.org/10.5281/Z

ENODO.6053272

Saranyaraj, D., Manikandan, M., & Maheswari, S. (2020).

A deep convolutional neural network for the early

detection of breast carcinoma with respect to hyper-

parameter tuning. Multimedia Tools and Applications,

79(15–16), 11013–11038. https://doi.org/10.1007/S110

42-018-6560-X/TABLES/12

Shen, L., Margolies, L. R., Rothstein, J. H., Fluder, E.,

McBride, R., & Sieh, W. (2019). Deep Learning to

Improve Breast Cancer Detection on Screening

Mammography. Scientific Reports, 9(1). https://

doi.org/10.1038/S41598-019-48995-4

The American Cancer Society medical and editorial content

team. (2022, January 12). Key Statistics for Breast

Cancer.

Applied Deep Learning Architectures for Breast Cancer Screening Classiﬁcation

623

Waskom, M. (2021). seaborn: statistical data visualization.

Journal of Open Source Software, 6(60), 3021.

https://doi.org/10.21105/JOSS.03021

Worsley, K. J. (1977). A Non-Parametric Extension of a

Cluster Analysis Method by Scott and Knott. Biometrics,

33(3), 532. https://doi.org/10.2307/2529369

Zerouaoui, H., & Idri, A. (2022). Deep hybrid architectures

for binary classification of medical breast cancer images.

Biomedical Signal Processing and Control, 71, 103226.

https://doi.org/10.1016/J.BSPC.2021.103226

Zerouaoui, H., Idri, A., Nakach, F. Z., & Hadri, R. el.

(2021). Breast Fine Needle Cytological Classification

Using Deep Hybrid Architectures. 186–202. https://

doi.org/10.1007/978-3-030-86960-1_14

Zhang, Z. (2019). Improved Adam Optimizer for Deep

Neural Networks. 2018 IEEE/ACM 26th International

Symposium on Quality of Service, IWQoS 2018.

https://doi.org/10.1109/IWQOS.2018.8624183

Zizaan, A., & Idri, A. (2022). Systematic Literature review

of Machine Learning based Breast Cancer Screening.

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

624