ResNet-101 Empowered Deep Learning for Breast Cancer Ultrasound

Image Classiﬁcation

Agnesh Chandra Yadav

1 a

, Maheshkumar H. Kolekar

1 b

and Mukesh Kumar Zope

Department of Electrical Engineering, IIT Patna, Patna, India

Department Medical Physics, IGIMS Patna, Patna, India

Keywords:

Breast Cancer Classiﬁcation, Ultrasound Images, Deep Learning, ResNet-101.

Abstract:

In the modern era, accurate breast cancer classiﬁcation plays a crucial role in early detection and treatment

planning. This article introduces a modiﬁed ResNet-101 architecture tailored speciﬁcally for classifying breast

cancer using ultrasound images. The ultrasound images undergo pre-processing before passing through our

adapted ResNet-101 model, which includes the integration of shortcut connections to enhance gradient stabil-

ity and deep structure adaptability for effective learning and classiﬁcation. The dataset comprises 780 images

categorized into normal, benign, and malignant cases. To address class imbalance, data augmentation tech-

niques are employed, enriching diversity and enhancing modeling precision. The proposed model achieves

exceptional performance, boasting precision, recall, F1-score, and accuracy values of 0.9855, 0.9677, 0.9756,

and 0.9743, respectively. The comparative analysis highlights the superiority of our model over existing tech-

niques. Furthermore, we explore its potential for clinical application using real-world datasets. Our ﬁndings

indicate signiﬁcant promise in revolutionizing breast cancer detection, offering a robust tool for early and ac-

curate diagnosis with the potential to impact patient outcomes greatly.

1 INTRODUCTION

As per the American Cancer Society’s projections

for 2022, there is an anticipated surge of approx-

imately 1,918,030 new cancer cases, leading to an

estimated 609,360 deaths within the United States

alone(Ferlay et al., 2018), (Loizidou et al., 2023).

Among these, breast cancer, an exceptionally preva-

lent and potentially life-threatening ailment affecting

women on a global scale, has risen to become the pri-

mary cause of mortality in nearly every nation. This

multifaceted disease, which accounts for about 30%

of all female cancers, necessitates timely identiﬁca-

tion and detection, as underscored by recent stud-

ies, to facilitate effective treatment and improve pa-

tient outcomes(Aavula et al., 2019)-(Chaurasia et al.,

2018). Cancer progresses through discernible stages,

and detecting it in an advanced phase presents con-

siderable risks. In the realm of medical image analy-

sis, particularly in the realm of breast cancer diagno-

sis, deep learning techniques have demonstrated sub-

stantial promise in enabling the accurate identiﬁcation

and classiﬁcation of breast cancer(Kaushik and Kaur,

2016). Over the past years, deep learning, a subset

https://orcid.org/0009-0007-8707-5142

https://orcid.org/0000-0002-4272-3528

of artiﬁcial intelligence, has emerged as a highly aus-

picious methodology in various medical domains, in-

cluding the detection of breast cancer(Rabiei et al.,

2022). Consequently, according to the World Health

Organization’s report in 2019, precise and early detec-

tion plays a pivotal role in advancing diagnosis and el-

evating the survival rate of breast cancer patients from

20% to 60%. With approximately 1.5 million women

receiving a diagnosis each year and half a million suc-

cumbing to the disease, breast cancer stands as a sig-

niﬁcant health challenge(Lotter et al., 2021).

Furthermore, the application of deep learning

methodologies in breast cancer detection holds the

promise of advancing personalized medicine through

its ability to offer insights into the subtype classiﬁca-

tion of breast cancer. This information stands as a piv-

otal factor in tailoring treatment regimens to individ-

ual patients, ultimately resulting in more precise ther-

apeutic interventions and enhanced prognostic out-

comes. Given the recent strides and encouraging out-

comes witnessed in the realm of deep learning-driven

breast cancer detection, there is a burgeoning impe-

tus among researchers to delve deeper into reﬁning

and expanding these methodologies. Through the ju-

dicious utilization of extensive datasets, ﬁne-tuning of

network architectures, and integration of multi-modal

imaging modalities, deep learning stands poised to re-

Yadav, A., Kolekar, M. and Zope, M.

ResNet-101 Empowered Deep Learning for Breast Cancer Ultrasound Image Classiﬁcation.

DOI: 10.5220/0012377800003657

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 1, pages 763-769

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

763

deﬁne the landscape of breast cancer diagnosis, im-

parting a substantial inﬂuence on patient care. This

impetus is substantiated by recent scholarly works,

including those by Ramadan et al. (Ramadan et al.,

2020), Hakin et al. (Hakim et al., 2021), Zeiser et

al. (Zeiser et al., 2020), and Rehman et al. (Rehman

et al., 2021), who have proffered CNN-based models

for breast cancer classiﬁcation. These investigations

underscore the vast potential of deep learning in the

realm of breast cancer detection and categorization,

laying a robust groundwork for continued exploration

and progress in this pivotal domain of research.

J. W. Li et al. (Li et al., 2022) undertook an in-

depth analysis of ultrasound images aimed at predict-

ing the behavior of breast invasive ductal carcinoma.

This endeavor offers a noninvasive means of assess-

ing, quantifying tumor characteristics, and tailoring

treatment decisions on an individualized basis. How-

ever, it is imperative to acknowledge and surmount

challenges such as data variability, optimal feature se-

lection, mitigating overﬁtting, conducting robust ex-

ternal validation, and ensuring clinical applicability

to facilitate the successful translation of these ﬁndings

into clinical practice.

In a separate study, Sharma et al. (Sharma et al.,

2020) engaged in a comprehensive multi-class classi-

ﬁcation analysis to compare the performance of di-

verse classiﬁers, including decision trees, k-nearest

neighbor (kNN), support vector machine (SVM), and

ensemble classiﬁers. Their primary objective was

the early prediction and detection of dementia. No-

tably, it is crucial to highlight that these classiﬁers

were not speciﬁcally employed for early breast can-

cer detection in their research. The focus of their in-

vestigation was squarely on dementia prediction, and

the aforementioned classiﬁers were rigorously eval-

uated within the conﬁnes of that speciﬁc context.

Aboutalib et al. (Aboutalib et al., 2018) conducted

a study delving into the application of groundbreak-

ing deep learning techniques to discern recalled yet

benign mammography images from negative exam-

inations and those displaying malignancy. Through

this approach, they achieved remarkable outcomes in

terms of accuracy, sensitivity, and speciﬁcity. By har-

nessing the potential of deep learning, the authors

showcase the possibility of enhancing the early de-

tection of breast cancer by precisely identifying and

distinguishing among various types of mammography

images. On a related note, S. Mishra et al. (Misra

et al., 2021) proposed an ensemble transfer learning

methodology that incorporates elastography and B-

mode breast ultrasound images, aiming to enhance di-

agnostic precision and generalization capabilities.

In 2022, several methodologies were developed

to advance breast cancer detection and classiﬁcation.

(Ueda et al., 2022) Developed and validated a deep

learning model for breast cancer detection in mam-

mography and a clinical decision support system us-

ing ultrasound images, with a limitation in obtain-

ing a diverse annotated dataset for robust training

and generalization. (Ragab et al., 2022) Introduced

an Ensemble Deep-Learning-Enabled Clinical Deci-

sion Support System utilizing VGG-16, VGG-19, and

SqueezeNet for feature extraction from ultrasound

images, highlighting enhanced performance but rais-

ing concerns about the computational resources re-

quired. (Jabeen et al., 2022) Proposed a breast cancer

classiﬁcation framework involving modiﬁcation of a

DarkNet-53 model, transfer learning, and optimiza-

tion algorithms, with potential challenges in the in-

terpretability of selected features. (Althobaiti et al.,

2022) Presented a deep transfer learning-based model

for breast cancer detection using photoacoustic mul-

timodal imaging, demonstrating promise but raising

questions about resilience to variations in imaging

conditions. (Jabeen et al., 2022) introduced an auto-

mated model for breast cancer diagnosis using digital

mammograms, incorporating pre-processing and hy-

perparameter tuning, with potential sensitivity to tun-

ing processes. However, it should be noted that this

approach comes with escalated computational com-

plexity and entails challenges in determining suitable

combination strategies and ensuring model diversity.

The aim of this current study was to employ a

range of varied deep-learning methodologies and in-

tegrate multiple inﬂuencing factors into the modeling

process for the prediction of breast cancer. The fol-

lowing key contributions underscore the signiﬁcance

of this work:

1. We have modiﬁed the ResNet-101 architecture

speciﬁcally for breast cancer ultrasound image

classiﬁcation. In this model, we have employed

Shortcut Connections for Gradient Stability to ad-

dress the vanishing gradient problem with ’Bot-

tleneck’ blocks.

2. In our proposed model, we have implemented a

mechanism for Deep Structure Adaptability, al-

lowing the network to effectively handle intricate

architectures, facilitating enhanced learning and

classiﬁcation capabilities.

3. We have conducted a comparative analysis be-

tween the proposed modiﬁed ResNet-101 model

and various ResNet base models.

4. We further evaluate our proposed model alongside

state-of-the-art methods to validate its reliability

and effectiveness.

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

764

2 MATERIAL AND METHODS

Fig. 1 presents a schematic representation of the

methodology employed in this research. For a com-

prehensive understanding of each method and its

practical execution, please refer to the respective sec-

tions, namely Sections 2.1 to 2.3, where you can ﬁnd

in-depth explanations and implementation speciﬁcs.

Figure 1: Block diagram of the proposed framework.

2.1 Data Collection

At the initial data collection phase, breast ultrasound

images were acquired from a cohort of female sub-

jects spanning an age range of 25 to 75 years. This

data acquisition process took place in the year 2018.

The total patient count stands at 600 individuals, all

of whom are female. The dataset encompasses a total

of 780 images, with each image possessing an aver-

age dimension of 500 pixels in both width and height

(500x500 pixels). The images are formatted in PNG

(Portable Network Graphics) format. Each original

image is accompanied by its corresponding ground

truth image. These images are systematically classi-

ﬁed into three distinct classes, namely ’normal’, ’be-

nign’, and ’malignant’, facilitating the categorization

and analysis of breast conditions(Al-Dhabyani et al.,

2020). Exemplar samples from the breast ultrasound

dataset are presented in Fig. 2 for reference.

Figure 2: Few examples of dataset.

2.2 Data Preprocessing

The study incorporates a dataset of over 780 distinct

images, allocated for training, validation, and testing

in ratios of approximately 72: 13: 15, respectively.

Speciﬁcally, the training set consists of 563 images,

with 315 classiﬁed as benign, 152 as malignant, and

96 as normal. The validation set comprises 100 im-

ages, distributed as 56 benign, 27 malignant, and 17

normal cases. Additionally, the testing set encom-

passes 117 images, with 66 benign, 31 malignant, and

20 normal instances. To ensure precise evaluation,

a new folder was meticulously established for test-

ing due to the absence of a predeﬁned testing folder.

In order to address the inherent imbalanced distribu-

tion within each group, data augmentation techniques

were systematically applied using an Image Gener-

ator. This approach was strategically employed to

enrich the dataset’s diversity and effectively mitigate

the challenge of class imbalance, thereby fostering a

more resilient and precise modeling for breast can-

cer prediction. Diverse data augmentation techniques

were tactically employed to accomplish this objective

2.3 Enhancing Breast Cancer

Prognostication Through Advanced

Predictive Modeling

In this section, we will elucidate our devised pre-

dictive modeling method for the classiﬁcation of

breast cancer as shown in Fig. 3. In our proposed

Figure 3: Our proposed modiﬁed ResNet-101 architecture.

model, we have made signiﬁcant modiﬁcations to the

ResNet-101 architecture tailored speciﬁcally for clas-

sifying breast cancer ultrasound images. This adapted

ResNet-101 structure comprises multiple layers, each

carrying out a sequence of operations. To start, the

initial layer, denoted as ’conv1’, applies a 2D convo-

lution operation to the input with 3 input channels and

generates 64 output channels. It employs a kernel size

of 7x7 and a stride of 2. Following this, batch nor-

ResNet-101 Empowered Deep Learning for Breast Cancer Ultrasound Image Classiﬁcation

765

malization (’bn1’) is applied, which standardizes the

activations, and a rectiﬁed linear unit (ReLU) activa-

tion function is used to introduce non-linearity. Sub-

sequently, the ’max pool’ layer conducts max pooling

with a kernel size of 3x3 and a stride of 2, reducing the

spatial dimensions of the data. The subsequent lay-

ers, namely ’layer1’, ’layer2’, ’layer3’, and ’layer4’,

are constructed as stacks of ’Bottleneck’ blocks. Each

of these blocks encompasses a series of convolutional

layers, combined with batch normalization and ReLU

activation functions. Additionally, they feature a dis-

tinctive ”shortcut” connection that allows information

to bypass certain layers. This is instrumental in miti-

gating the vanishing gradient problem, a common is-

sue in deep neural networks. The number of output

channels in each ’Bottleneck’ block varies, gradually

increasing, which enables the network to capture pro-

gressively complex features. Importantly, the archi-

tecture is purposefully designed to handle very deep

structures while remaining trainable, which is a cru-

cial factor in achieving effective learning and classi-

ﬁcation performance. The model employs the Cat-

egorical Cross Entropy loss function and undergoes

optimization via an Adam optimizer with a learning

rate set at 0.0001. The training procedure spans 20

epochs, employing a batch size of 8.

3 RESULTS AND DISCUSSIONS

To apply our model to the dataset by Dhabyani et

al.(Al-Dhabyani et al., 2020), we partitioned the data

into three categories: Benign tumor cell, Malignant

tumor cell, and Normal. The study encompasses more

than 563 unique images, allocating 100 for testing,

117 for validation, and 780 for training. The ratios

for training, validation, and testing were conﬁgured

at 72:13:15. The entire experiment was conducted

utilizing Google Colab and Jupyter Notebook. We

performed multiple model simulations to assess the

performance of our proposed system.

We evaluated several performance metrics to vali-

date our proposed modiﬁed ResNet-50 model, includ-

ing accuracy (Acc), precision (Pre), sensitivity (Sen)

or recall (Rec), and F1-score. These metrics were cal-

culated using the following formulas:

Acc =

T P + T N

T P + FP + FN + T N

(1)

Rec = Sen =

T P

T P + FN

(2)

Pre =

T P

FP + T P

(3)

F − measure =

2 ∗ Pre ∗ Rec

Pre + Rec

(4)

In our study, we conducted an evaluation of var-

ious parameters on each group of datasets, namely

Benign, Malignant, and Normal, as outlined in Ta-

ble 1 and this table provides precise performance met-

rics for our breast cancer classiﬁcation model across

distinct groups. For ”Benign tumors,” the model

achieves a notable precision of 0.9565, indicating that

approximately 0.9565 of predicted benign cases were

accurately classiﬁed. Furthermore, the recall score of

1.0000 implies that all actual benign cases were suc-

cessfully identiﬁed. The F1-score, which balances

precision and recall, attains an impressive 0.9777.

Shifting to the ”Normal” category, both precision and

recall stand at a perfect 1.0000, signifying ﬂawless

classiﬁcation. In the realm of ”Malignant tumors,” the

model demonstrates impeccable precision at 1.0000,

afﬁrming that all predicted malignant cases were in-

deed correct. However, the recall rate of 0.9032 sug-

gests that there were a few actual malignant cases

that the model missed. The F1-score in this category

amounts to 0.9491. The ”Overall average” metrics

amalgamate these assessments, yielding an average

precision of 0.9855, a recall of 0.9677, and an F1-

score of 0.9756. These aggregated metrics present a

comprehensive evaluation of the model’s proﬁciency

across all classiﬁcation groups, portraying a robust

performance in the classiﬁcation of breast cancer.

Table 1: Performance matrices.

Parameters/Group Precision Recall F1-score

Benign tumor 0.9565 1.0000 0.9777

Normal 1.0000 1.0000 1.0000

Malignant tumor 1.0000 0.9032 0.9491

Overall average 0.9855 0.9677 0.9756

In addition, we assessed testing and validation ac-

curacy using a dataset consisting of 100 images. The

results demonstrated a commendable accuracy level,

with both types of breast cancer groups achieving a

score of 0.9743. This indicates a robust performance

in accurately classifying cases within the dataset. Fig.

4 displays the predicted results obtained from our pro-

posed modiﬁed ResNet-101 model. Fig. 5 shown the

confusion matrix of our proposed model.

Table 2: Previous methods compared with our proposed

model.

Author Method Pre Rec F1-score Acc

Ramadan et al. (Ramadan et al., 2020) CNN - 0.9141 - 0.9210

Hakin et al. (Hakim et al., 2021) CNN - 0.8812 - 0.9034

Fanizzi et al. (Fanizzi et al., 2020) Random Forest - 0.8910 - 0.8855

Rehman et al. (Rehman et al., 2021) Fully connected CNN - 0.9700 - 0.8701

Zeiser et al. (Zeiser et al., 2020) CNN (U-NET) - 0.9230 - 0.8590

R. Rabiei et al. (Rabiei et al., 2022) Random Forest - 0.9304 - 0.8004

Proposed Method CNN (Resnet-101) 0.9855 0.9677 0.9756 0.9756

In Table 2, we present a comprehensive compar-

ative evaluation of various breast cancer classiﬁca-

tion methodologies. Each approach was rigorously

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

766

Figure 4: Few predicted samples of our proposed model.

Figure 5: Confusion matrix of our proposed model.

examined for its capacity to accurately discern breast

cancer cases. The assessed methods, along with their

corresponding performance metrics, are meticulously

detailed. Ramadan et al. harnessed the power of a

CNN and attained a precision of 0.9141 and an ac-

curacy of 0.9210 (Ramadan et al., 2020). Hakin et

al. similarly employed a CNN, yielding a precision

of 0.8812 and an accuracy of 0.9034 (Hakim et al.,

2021). Fanizzi et al. opted for a Random Forest

technique, yielding a precision of 0.8910 and an ac-

curacy of 0.8855 (Fanizzi et al., 2020). Rehman et

al. implemented a Fully Connected CNN, achiev-

ing a remarkable precision of 0.9700, albeit with a

marginally lower accuracy of 0.8701 (Rehman et al.,

2021). Zeiser et al. employed a CNN with U-NET ar-

chitecture, securing a precision of 0.9230 and an ac-

curacy of 0.8590 (Zeiser et al., 2020). R. Rabiei et

al. utilized a Random Forest approach, resulting in a

precision of 0.9304 and an accuracy of 0.8004 (Rabiei

et al., 2022). However, our proposed method, leverag-

ing a CNN with ResNet-101 architecture, unequivo-

cally outperforms all alternative techniques. It attains

an exceptional precision of 0.9855, underscoring an

exceedingly low false positive rate. This precision

is of paramount importance in the medical domain,

where precise identiﬁcation of positive cases holds ut-

most signiﬁcance. Furthermore, the model exhibits

a commendable recall score of 0.9677, signifying its

adeptness in capturing all bonaﬁde positive cases and

thereby minimizing false negatives. The F1-score, a

pivotal metric that strikes a balance between preci-

sion and recall, reaches an impressive 0.9756, further

corroborating the model’s robustness. The overall ac-

curacy, standing at 0.9756, unequivocally establishes

the proposed method’s superior performance. While

the other methods demonstrate commendable perfor-

mance, none surpasses the precision, recall, and F1-

score achieved by our proposed CNN with ResNet-

101 architecture.

Overall, our proposed breast cancer classiﬁcation

model demonstrates superior performance compared

to the random forest-based approaches presented by

R. Rabiei et al. (Rabiei et al., 2022) and Fanizzi et

al. (Fanizzi et al., 2020) across key metrics includ-

ing accuracy, recall, precision, and F1 score. Fur-

thermore, our model achieves a competitive accuracy

when compared to the fully connected depth-wise

separable CNN model introduced by Rehman et al.

(Rehman et al., 2021), while also attaining higher re-

call, precision, and F1-score. These ﬁndings under-

score the efﬁcacy of our model in the precise clas-

siﬁcation of breast cancer utilizing breast ultrasound

images.

3.1 Ablation Study

In this comprehensive ablation study, we systemat-

ically assessed the performance of various ResNet

architectures, as detailed in Table 3. The models

evaluated include ResNet-18, ResNet-34, ResNet-50,

ResNet-152, and ResNet-101, alongside a proposed

modiﬁed ResNet-101 speciﬁcally tailored for breast

cancer classiﬁcation and segmentation tasks. The

baseline models demonstrated incremental improve-

ments in accuracy, precision, recall, and F1-score

with increasing depth, with ResNet-152 achieving

the highest accuracy. Notably, ResNet-101 exhibited

competitive accuracy and an impressive F1-score of

0.9760. However, the proposed modiﬁed ResNet-101

surpassed all models, showcasing its prowess with the

highest precision and recall at 0.9855, underscoring

its ability to capture nuanced features crucial for accu-

rate breast cancer classiﬁcation, as evidenced in Table

ResNet-101 Empowered Deep Learning for Breast Cancer Ultrasound Image Classiﬁcation

767

Table 3: Performance metrics of diverse ResNet architec-

tures and the proposed modiﬁed ResNet-101.

Model/Parameters Pre Rec F1-score Acc

ResNet-18 0.8921 0.9123 0.9014 0.8898

ResNet-34 0.9040 0.9154 0.9102 0.9001

ResNet-50 0.9308 94.02 0.9405 0.9274

ResNet-152 0.9621 0.9614 0.9652 0.9847

ResNet-101 0.9542 0.9548 0.9760 0.9511

Proposed Mod-ResNet-101 0.9855 0.9855 0.9756 0.9756

Incorporating the segmentation task into the eval-

uation, the proposed modiﬁed ResNet-101 demon-

strated its versatility, as depicted in Figure 4. Seg-

mentation demands a nuanced understanding of im-

age features, and the model’s modiﬁcations evidently

contributed to its efﬁcacy in delineating regions of in-

terest. The precision and recall metrics in the seg-

mentation task aligned with those of the classiﬁcation

task, highlighting the model’s consistent performance

across both domains. This dual-task capability proves

instrumental in providing a holistic solution for breast

cancer analysis, where accurate localization and clas-

siﬁcation of abnormalities are paramount. While op-

portunities for ﬁne-tuning and further improvement

exist, this study ﬁrmly establishes the proposed mod-

iﬁed ResNet-101 as a robust and versatile architec-

ture for comprehensive breast cancer analysis, effec-

tively addressing both classiﬁcation and segmenta-

tion tasks. By offering a comprehensive solution for

accurate localization and classiﬁcation of abnormali-

ties, our meticulously designed architecture stands as

a beacon of innovation, paving the way for heightened

precision in medical image analysis and holding im-

mense promise for improving diagnostic outcomes in

breast cancer research and clinical practice.

4 CONCLUSION

Our proposed modiﬁed ResNet-101 architecture for

breast cancer classiﬁcation via breast ultrasound im-

agery has demonstrated superior performance com-

pared to existing approaches. It exhibited height-

ened accuracy, recall, precision, and F1-score met-

rics, signifying its efﬁcacy in precisely discerning

cases of breast cancer. These outcomes outperformed

random forest-based models and rivaled a fully con-

nected depth-wise separable CNN model. These re-

sults underscore the potential of deep learning archi-

tectures in breast cancer classiﬁcation and provide a

solid groundwork for future investigations. Through

continued reﬁnement and progression of these mod-

els, we can make signiﬁcant strides in early and pre-

cise breast cancer identiﬁcation, ultimately leading to

enhanced patient outcomes and reduced disease bur-

den.

Moving forward, we have outlined several areas

for potential improvement in breast cancer classiﬁ-

cation models. A pivotal aspect of our agenda is to

leverage a distinct dataset sourced from our institution

(IGIMS Patna) to validate our model using authentic

clinical data. This endeavor will allow us to evalu-

ate the model’s performance and resilience in a real-

world healthcare setting, offering further substanti-

ation of its adeptness in accurately detecting breast

cancer.

REFERENCES

Aavula, R., Bhramaramba, R., and Ramula, U. S. (2019). A

comprehensive study on data mining techniques used

in bioinformatics for breast cancer prognosis. Journal

of Innovation in Computer Science and Engineering,

9(1):34–39.

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

M. L., Sumkin, J. H., and Wu, S. (2018). Deep learn-

ing to distinguish recalled but benign mammography

images in breast cancer screening. Clinical Cancer

Research, 24(23):5902–5909.

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

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

brief, 28:104863.

Althobaiti, M. M., Ashour, A. A., Alhindi, N. A., Al-

thobaiti, A., Mansour, R. F., Gupta, D., Khanna, A.,

et al. (2022). Deep transfer learning-based breast can-

cer detection and classiﬁcation model using photoa-

coustic multimodal images. BioMed Research Inter-

national, 2022.

Chaurasia, V., Pal, S., and Tiwari, B. (2018). Prediction of

benign and malignant breast cancer using data mining

techniques. Journal of Algorithms & Computational

Technology, 12(2):119–126.

Fanizzi, A., Basile, T. M., Losurdo, L., Bellotti, R., Bottigli,

U., Dentamaro, R., Didonna, V., Fausto, A., Massafra,

R., Moschetta, M., et al. (2020). A machine learning

approach on multiscale texture analysis for breast mi-

crocalciﬁcation diagnosis. BMC bioinformatics, 21:1–

11.

Ferlay, J., Ervik, M., Lam, F., Colombet, M., Mery, L.,

neros, M., Znaor, A., Soerjomataram, I., and Bray,

F. (2018). Global cancer observatory: cancer today.

Lyon, France: international agency for research on

cancer, 3(20):2019.

Hakim, A., Prajitno, P., and Soejoko, D. (2021). Micro-

calciﬁcation detection in mammography image using

computer-aided detection based on convolutional neu-

ral network. In AIP Conference Proceedings, volume

2346. AIP Publishing.

Jabeen, K., Khan, M. A., Alhaisoni, M., Tariq, U., Zhang,

Y.-D., Hamza, A., Mickus, A., and Dama

sevi

cius, R.

(2022). Breast cancer classiﬁcation from ultrasound

images using probability-based optimal deep learning

feature fusion. Sensors, 22(3):807.

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

768

Kaushik, D. and Kaur, K. (2016). Application of data

mining for high accuracy prediction of breast tissue

biopsy results. In 2016 Third International Confer-

ence on Digital Information Processing, Data Mining,

and Wireless Communications (DIPDMWC), pages

40–45. IEEE.

Li, J.-w., Cao, Y.-c., Zhao, Z.-j., Shi, Z.-t., Duan, X.-

q., Chang, C., and Chen, J.-g. (2022). Prediction

for pathological and immunohistochemical character-

istics of triple-negative invasive breast carcinomas:

the performance comparison between quantitative and

qualitative sonographic feature analysis. European

Radiology, pages 1–11.

Loizidou, K., Elia, R., and Pitris, C. (2023). Computer-

aided breast cancer detection and classiﬁcation in

mammography: A comprehensive review. Computers

in Biology and Medicine, page 106554.

Lotter, W., Diab, A. R., Haslam, B., Kim, J. G., Grisot, G.,

Wu, E., Wu, K., Onieva, J. O., Boyer, Y., Boxerman,

J. L., et al. (2021). Robust breast cancer detection in

mammography and digital breast tomosynthesis using

an annotation-efﬁcient deep learning approach. Na-

ture Medicine, 27(2):244–249.

Misra, S., Jeon, S., Managuli, R., Lee, S., Kim, G., Lee, S.,

Barr, R. G., and Kim, C. (2021). Ensemble transfer

learning of elastography and b-mode breast ultrasound

images. arXiv preprint arXiv:2102.08567.

Rabiei, R., Ayyoubzadeh, S. M., Sohrabei, S., Esmaeili, M.,

and Atashi, A. (2022). Prediction of breast cancer us-

ing machine learning approaches. Journal of Biomed-

ical Physics & Engineering, 12(3):297.

Ragab, M., Albukhari, A., Alyami, J., and Mansour, R. F.

(2022). Ensemble deep-learning-enabled clinical de-

cision support system for breast cancer diagnosis

and classiﬁcation on ultrasound images. Biology,

11(3):439.

Ramadan, S. Z. et al. (2020). Using convolutional neural

network with cheat sheet and data augmentation to de-

tect breast cancer in mammograms. Computational

and Mathematical Methods in Medicine, 2020.

Rehman, K. U., Li, J., Pei, Y., Yasin, A., Ali, S., and Mah-

mood, T. (2021). Computer vision-based microcalci-

ﬁcation detection in digital mammograms using fully

connected depthwise separable convolutional neural

network. Sensors, 21(14):4854.

Sharma, N., Kolekar, M. H., and Jha, K. (2020). Iterative

ﬁltering decomposition based early dementia diagno-

sis using eeg with cognitive tests. IEEE Transactions

on Neural Systems and Rehabilitation Engineering,

28(9):1890–1898.

Ueda, D., Yamamoto, A., Onoda, N., Takashima, T., Noda,

S., Kashiwagi, S., Morisaki, T., Fukumoto, S., Shiba,

M., Morimura, M., et al. (2022). Development

and validation of a deep learning model for detec-

tion of breast cancers in mammography from multi-

institutional datasets. PLoS One, 17(3):e0265751.

Zeiser, F. A., da Costa, C. A., Zonta, T., Marques, N. M.,

Roehe, A. V., Moreno, M., and da Rosa Righi, R.

(2020). Segmentation of masses on mammograms us-

ing data augmentation and deep learning. Journal of

digital imaging, 33:858–868.

ResNet-101 Empowered Deep Learning for Breast Cancer Ultrasound Image Classiﬁcation

769