Prediction and Early Diagnosis of Breast

Cancer Using Deep Learning

Hadeer Saber Mohamed Yousry Naglaa Fathy Tamer Abdelkader

Dept. of Computer Engineering Information Systems Information Systems Dept. Faculty of Computer Science
The British University in Egypt Dept. Faculty of Computer and and Engineering
Cairo, Egypt Faculty of Computer and Information Sciences Galala University
[email protected] Information Sciences Ain Shams University [email protected]
Ain Shams University Cairo, Egypt
[email protected]. [email protected]
edu.eg

Abstract— Breast cancer is the most prevalent type of cancer The gold standard for diagnosis is histopathological
and the leading cause of cancer-related fatalities globally. Early analysis of biopsy tissues, which helps determine whether a
detection is crucial but often difficult, resulting in severe tumor is benign or malignant through microscopic
consequences for many individuals. This work employs deep examination and staining techniques like H&E[3].
learning, particularly the VGG-16 Convolutional Neural
Network (CNN) framework, to identify and diagnose breast A. Motivation
cancer in histopathology images. A comprehensive breast Using deep learning and machine learning to diagnose
cancer image dataset was examined using Random Forest, breast cancer seeks to enhance patient outcomes, lessen the
ResNet-50, VGG-16, and a custom CNN model. The VGG-16 disease's worldwide impact, and promote improvements in
model attained an accuracy of 86.96% in distinguishing between medical research and technology.
benign and malignant tissues. The ResNet-50 and CNN models,
combined with Random Forest classifiers, achieved accuracies B. Benefits of this Research
of 79.99%, 95.95%, and 99.93%, respectively, with the top
The benefits of this research are as follows; it enhances
model reaching a recall of 99% and a precision of 99.88%. This
tailored therapy and early diagnosis. Furthermore, it possesses
indicates that ensemble learning techniques can sometimes
surpass traditional deep learning methods. The research also the capacity to revolutionize the treatment of breast cancer and
highlights the significance of tumor markers such as HER2, ER, preserve lives by impacting domains like investigation,
PR, and Ki-67 in determining malignancy, helping to assess financial viability, and global availability of medical services.
tumor aggressiveness and potential treatment responses. This The capabilities of machine learning in breast cancer
study demonstrates that integrating deep learning and machine diagnosis allow for lower healthcare expenses and a more
learning models can enhance the precision and clinical relevance positive impact on the management of breast cancer diagnosis.
of breast cancer diagnosis, thereby potentially reducing the
global impact of the disease and saving lives. II. SUMMARY OF PREVIOUS WORK
Now, interestingly, breast cancer detection has moved
Keywords—Breast Cancer, Random Forest, Deep Learning, forward with deep learning along with machine learning
Medical images, early diagnosis techniques. Deep learning complemented traditional machine
learning techniques to achieve better results in all functional
I. INTRODUCTION
parameters - accuracy and efficiency.
Breast cancer is a significant global health issue, ranking
as the second most frequent cancer after lung cancer and the Recent research has indicated various innovations. In one,
fifth leading cause of cancer-related deaths. Annually, it various like Google Net, VGG11, and MobileNetV3_Small
affects approximately 200,000 women and 1,700 men, were ensemble for pathology image classification of breast
causing around 40,000 and 450 deaths, respectively. In the tissue, and better performance was achieved by learning on
U.S., forecasts predict 207,090 new cases and 39,840 female datasets like BreakHis and ICIAR-2018. In another [4],
deaths. Among women aged 30-55, breast cancer is the VGG16 was used in mammography. High accuracy, 75.46%,
leading cause of death[1]. was recorded despite overfitting and computational problems
associated with it.
Detection typically begins with mammography, followed
by computer-aided image analysis to identify suspicious areas. For instance, a combination of convolutional neural
Key tumor markers in histopathology images, such as HER2, networks with image segmentation has shown tremendous
ER, PR, and Ki-67 [2], play a crucial role in diagnosis and potential so far while obtaining a minimum accuracy of 84%;
treatment planning. HER2 indicates aggressive cancer, while therefore, in the classification of mammogram images that
ER and PR suggest hormone receptor-positive tumors will help health experts quickly diagnose, further potential
responsive to hormonal therapies. High Ki-67 levels denote improvement could be achieved by a combination of models
rapidly proliferating aggressive cancer. with each other or tuning hyperparameters as shown by
improvement of binary classification accuracy up to 97.26% from mammographic images, capturing high-level features
through the application of the triple-concatenated model[5]. that may include traditional methods. This automated feature
extraction significantly enhances diagnostic accuracy, which
Transfer learning has been applied, for instance, to is important for clinical decision-making and early
classifying images of breast cancers, achieving good results; intervention. By combining these deep learning models with
enhanced image magnification and data augmentation traditional machine learning classifiers, the proposed
techniques have led to an accuracy of 91% in one study[6], methodology aims to further refine the classification process,
while another approach achieved 98.91% due to data ensuring a higher degree of accuracy and reliability in
augmentation methods leading to increased diversity of the detecting breast cancer at earlier stages. The integration of
training data, thus improving generalization and reducing deep learning pattern recognition capabilities with the
overfitting[7]. robustness of traditional algorithms offers a powerful tool in
Consequently, the hybrid techniques that make use of both the ongoing fight against breast cancer, potentially improving
these CNNs and the Term Variance feature selection patient outcomes through more accurate diagnoses.
algorithm, along with multiclass support vector machines of Data preprocessing is meticulously performed on the
MSVMs[8], present high accuracy in classifying fake news, BreakHis dataset to optimize model performance. This
reaching 99% in some cases - means the selection of the
includes normalization to standardize pixel values across
features is used to enhance the performance of the applied images and data augmentation techniques to diversify the
model. training dataset and improve model robustness. The use of
The work on staging of breast cancer using CNN-based directory flow generation facilitates the seamless integration
models appeared very promising and achieved 90% accuracy, of preprocessed data into the training pipeline, ensuring
which points out that the pre-trained models are better in efficient model training and evaluation.
results as compared to training from scratch [9]. Further CNN
transfer learning with augmentation of available data
stipulated images gave a high accuracy of 96.91 in the
histology images [10].
Overall, the presented studies provide evidence that the
incorporation of machine learning and deep learning models
offers significant advances in breast cancer detection with
improved diagnostic accuracy for potential clinical
applications.
A. Discussion of findings
1) Strengths: The reviewed studies show that
combining deep learning with traditional machine learning
enhances diagnostic accuracy. Techniques like transfer
learning, data augmentation, and model ensembles contribute
to higher performance and robustness.
2) Weaknesses: Challenges such as overfitting and
computational complexity are common. Addressing these Figure 1 . Flow chart showing the proposed prediction methodolgy
issues requires careful model tuning and resource
optimization. The Custom Convolutional Neural Network (CNN) model
3) Comparison with Other Work: The discussed processes input mammographic images through multiple
layers designed to extract increasingly complex and high-level
methods generally outperform single-model approaches,
features, essential for accurate breast cancer detection. The
highlighting the importance of hybrid techniques and feature architecture of the custom CNN includes several key
selection. Performance metrics such as accuracy, precision, components, such as convolutional layers, which are
and recall are consistently high, demonstrating the efficacy of responsible for detecting local patterns like edges and textures
these advanced methods. in the images, and pooling layers, which reduce the spatial
dimensions while retaining the most important information.
III. PROPOSED PREDICTION METHODOLOGY These layers work together to form a robust feature map,
The methodology employed in this study, as illustrated in which is ultimately passed through fully connected layers for
the flowchart in Figure 1, focuses on advancing the early final decision-making. At the end of the network, a
detection and classification of breast cancer by utilizing a classification layer with a sigmoid activation function is
hybrid approach that combines both deep learning and employed, outputting a probability that indicates the presence
traditional machine learning techniques. This integrative or absence of breast cancer. The training process is guided by
method aims to gather the strengths of both paradigms, the binary cross-entropy loss function, which is well-suited for
ensuring robust and precise results. Central to this approach binary classification tasks like this one, where the goal is to
are advanced deep learning models, such as ResNet50 and distinguish between two possible outcomes. The model is
VGG16, which are pre-trained on publicly available datasets optimized using the Adam optimizer, a widely used
like ImageNet. These models are then fine-tuned specifically optimization algorithm in deep learning due to its adaptive
for breast cancer detection, allowing for a more targeted learning rate and efficient convergence properties. To prevent
application in this medical domain. ResNet50 and VGG16 are overfitting, where the model might perform well on training
renowned for their ability to automatically extract patterns data but poorly on unseen data, early stopping mechanisms are
employed. These mechanisms monitor the model’s Precision is the percentage of correctly classified elements
performance on validation data and halt training when for a given class and is given as Eq. (2).
improvements plateau, ensuring the model remains
generalizable. Precision = T P/ (T P + F P) (2)

In addition to the custom CNN, well-established pre- F1 Score is the harmonic mean between precision and
trained models such as VGG16 and ResNet50 are also fine- sensitivity and is given as Eq. (3).
tuned through transfer learning for breast cancer detection.
These models, both initially trained on the extensive F1 − score = 2T P/ (2T P + F P + F N) (3)
ImageNet dataset, are adapted to the specific task at hand.
VGG16 employs a layered approach where the early layers, Recall or Sensitivity is a rate of positive observations that
which capture general image features such as shapes and are correctly predicted as positive and is given as Eq. (4).
textures, are frozen to preserve valuable pre-trained
knowledge. Only the later layers are fine-tuned on the breast Recall = T P/ (T P + F N) (4)
cancer dataset, allowing the model to learn domain-specific
features. To enhance feature extraction and minimize the risk In summary, our methodology employs advanced deep
of overfitting, the VGG16 architecture incorporates global learning models like ResNet50 and VGG16, fine-tuned on
average pooling, which reduces the number of parameters, and medical datasets, to enhance breast cancer detection by
dropout regularization, which randomly disables neurons automatically extracting detailed features from
during training to improve model robustness. Similarly, mammographic images. This approach not only leverages the
ResNet50, with its deep architecture and unique residual high accuracy potential and efficient feature extraction
connections, is fine-tuned to leverage its exceptional ability to capabilities of deep learning but also integrates traditional
learn hierarchical representations of images. Residual machine learning techniques like Random Forest in an
connections help the model avoid the problem of vanishing ensemble setup to further improve diagnostic accuracy.
gradients in deep networks, making it particularly effective for However, challenges such as data dependency, computational
extracting high-level features. In the case of ResNet50, only intensity, interpretability issues, risk of overfitting, and
the final layers are modified for breast cancer classification, integration complexity must be carefully addressed to ensure
while the core of the network remains intact to retain the the robustness and applicability of our methodology in clinical
knowledge gained from ImageNet. Both models, like the settings.
custom CNN, are trained using binary cross-entropy loss and
optimized with the Adam optimizer. Additional techniques IV. EXPERIMENT RESULTS
such as learning rate scheduling, which adjusts the learning The performance of various machine learning models for
rate over time, and model checkpoints, which save the best breast cancer diagnosis was evaluated using a range of well-
performing models during training, further ensure robust established metrics, including accuracy, precision, recall, F1
performance and stability. score, and mean squared error (MSE). Each of these metrics
provides critical insights into different aspects of model
In addition to deep learning models, a hybrid approach that performance, offering a comprehensive understanding of how
combines the feature extraction power of VGG16 with the effectively the models classify breast tissue samples as either
classification strength of a traditional Random Forest benign or malignant. Accuracy gives a general sense of how
classifier is explored. In this ensemble method, features many samples were classified correctly, while precision
extracted from breast cancer images by VGG16 are fed into a measures the proportion of true positive results out of all
Random Forest algorithm for classification. The Random positive predictions, ensuring that false positives are
Forest, known for its robustness and ability to handle complex minimized. Recall, also known as sensitivity, gauges the
data distributions, complements the deep learning approach by model’s ability to identify actual positive cases, making it a
providing a strong, reliable classification mechanism. This crucial metric for medical diagnoses where missing a
combination enhances diagnostic accuracy, as it leverages the malignant case could have severe consequences. The F1 score
deep learning model’s ability to extract rich, high-dimensional provides a balance between precision and recall, offering a
features from images while utilizing the Random Forest’s single metric that accounts for both false positives and false
strength in decision-making and handling various feature negatives. Finally, mean squared error (MSE) reflects the
interactions. By merging these two approaches, this hybrid average squared difference between predicted and actual
method offers a powerful tool for improving the accuracy and outcomes, highlighting the overall discrepancy in the model’s
reliability of breast cancer detection, demonstrating how predictions.
traditional machine learning with cutting-edge deep learning
architectures can lead to better outcomes in medical Among the models tested, the Random Forest model
diagnostics. demonstrated exceptional performance across all the
evaluation metrics. With an almost perfect accuracy of
The algorithms are evaluated based on accuracy, 99.93%, the Random Forest algorithm consistently made
precision, recall, f1-score, and mean square error. Accuracy is correct predictions regarding the classification of breast tissue
the measure of the prediction made by algorithms and is given samples. This incredibly high accuracy indicates that the
as Eq. (1). model was able to effectively distinguish between benign and
malignant samples. The model’s precision was also
Accuracy = (T P + T N)/ (T N + F P + F N + T P) (1) impressive, scoring 99.88%, which means that nearly all of its
positive classifications were correct, with very few false
Where TP is true positive, TN is true negative, FP is false positives. This is particularly important in medical
positive, and FN is false negative. diagnostics, where a false positive could lead to unnecessary
anxiety and potentially harmful treatments for patients. The
model’s recall, at 99%, suggests it identified nearly all actual
malignant cases, ensuring that very few positive instances
were missed a crucial factor in early breast cancer detection.
The F1 score, calculated at 99.94%, demonstrates that the
model maintains a near-perfect balance between precision and
recall, excelling in both minimizing false positives and
maximizing the detection of true positives. Additionally, the
low mean squared error (MSE) of 0.0007 highlights the
model’s consistency and reliability, as it had minimal Figure 2. Analysis of Machine and Deep Learning Model Results
differences between predicted and actual outcomes.
In comparison, the Custom CNN model achieved a Overall, the Random Forest model emerged as the most
moderate accuracy of 79.99%, indicating that while it effective in this study, outperforming the other models across
performed reasonably well, there was still room for all evaluation metrics, including accuracy, precision, recall,
improvement in classifying breast cancer images. Its precision F1 score, and mean squared error (MSE). With its nearly
of 83.50% shows that a significant portion of its positive flawless accuracy and balanced performance in both precision
predictions were correct, although it was less reliable than the and recall, the Random Forest model demonstrated a superior
Random Forest model in terms of minimizing false positives. ability to classify breast tissue samples correctly, minimizing
The recall, at 85.62%, demonstrates the model's ability to both false positives and false negatives. Its exceptionally low
identify most of the actual positive cases, though it did miss MSE further emphasized the model's reliability, with very few
some malignant samples. The F1 score of 84.55% reflects a discrepancies between predicted and actual outcomes. This
fair balance between precision and recall, but it does suggest combination of factors makes the Random Forest model an
that the model had some inconsistencies in handling both ideal candidate for breast cancer diagnosis, particularly in
types of errors. The mean squared error (MSE) of 0.1465 clinical settings where precision and sensitivity are critical.
reveals a higher degree of discrepancy between predicted and The VGG16 model also delivered strong results,
actual outcomes, further indicating that the Custom CNN particularly in terms of precision, where it outperformed all
model's predictions were less accurate than those of the other models. Its high precision rate is crucial in minimizing
Random Forest model. false positive diagnoses, which can lead to unnecessary and
ResNet50, another model evaluated, demonstrated strong potentially harmful medical interventions. While its accuracy
performance with a high accuracy of 95.95%, showcasing its and recall were slightly lower than those of the Random Forest
capability to effectively classify breast cancer images. The model, VGG16 maintained a well-balanced performance
model achieved a precision of 84.02%, meaning that the across all metrics. The low MSE indicates that it made
majority of its positive predictions were accurate, although relatively few significant errors, highlighting its potential as a
there was still a small margin of false positives. The recall reliable diagnostic tool when optimized for breast cancer
score of 82.52% indicates that the model was able to detect detection through transfer learning.
most of the malignant cases, though it did miss a small On the other hand, the Custom CNN and ResNet50
number. The F1 score of 83.26% shows a reasonable balance models, while demonstrating moderate success, highlighted
between precision and recall, confirming that the model is the need for further refinement and optimization. The Custom
generally reliable, but there is room for improvement in CNN model, with its lower accuracy and higher MSE, showed
further optimizing the trade-off between these two metrics. that while it could effectively identify many positive cases,
The mean squared error (MSE) of 0.1582 suggests that, while there was a notable margin for improvement in reducing
ResNet50 performed well overall, there were still some prediction errors. This suggests that deeper or more complex
discrepancies between its predictions and the actual outcomes, architectures, additional data augmentation, or enhanced
indicating that further fine-tuning of the model may be regularization techniques could improve its performance.
necessary for optimal performance in this domain. Similarly, the ResNet50 model, despite leveraging its residual
The VGG16 model achieved a notable accuracy of 86.96%, connections and deep architecture, fell short in terms of
demonstrating that it correctly classified a large proportion of precision and accuracy compared to the top-performing
breast cancer images. Its precision of 91.89% was the highest models. While its F1 score demonstrated a balance between
precision and recall, the relatively higher MSE suggested
among all the models evaluated, indicating that nearly all
inconsistencies in its predictions. Further fine-tuning, such as
positive classifications (malignant cases) were accurate, with
adjusting learning rates, adding more training data, or
very few false positives. This makes VGG16 particularly exploring different transfer learning strategies, could help
strong in reducing unnecessary treatments for false positive boost its performance in future iterations.
diagnoses. The recall score of 87.31% suggests that the model
was highly effective at identifying the majority of malignant Confusion Matrices
cases, although it still missed some. The F1 score of 89.54% Confusion matrices provide a detailed view of each
indicates that the model strikes a good balance between model’s classification performance, showing the distribution
precision and recall, effectively managing both false positives of true positives, true negatives, false positives, and false
and false negatives. The model’s mean squared error (MSE) negatives.
of 0.1056 was relatively low, highlighting that it made few Confusion Matrix
True Label

significant errors in its predictions, contributing to its

reliability in breast cancer diagnosis. Benign 2400 0
 Malignant 0 4320 The visual analysis of the experimental results provides a
detailed understanding of the performance of different models
Benign Malignant in diagnosing breast cancer. The Random Forest model
consistently demonstrates superior performance across all
Prediction Label metrics, making it the most effective model in this study. The
VGG16 model also shows strong performance, particularly in
Figure 3: Confusion Matrix for Random Forest terms of precision, while the Custom CNN and ResNet50
models exhibit the need for further optimization to reduce
The confusion matrix for the Random Forest model, in errors and improve predictive accuracy.
Figure 3, shows zero false positives and false negatives,
indicating its high accuracy and reliability in classifying breast
tissue samples.
V. CONCLUSION & FUTURE WORK
Confusion Matrix This study conducted a comprehensive evaluation of
advanced machine learning and deep learning models for
Benign 426 174
detecting breast cancer using histopathological images,
True Label

focusing on a custom Convolutional Neural Network (CNN),

Malignant 146 918
ResNet-50, VGG-16, and Random Forest. The findings
revealed that the Random Forest model achieved the highest
Benign Malignant
performance, with an impressive 99.93% accuracy, 99.88%
precision, 99% recall, and a 99.94% F1 score. These results
Prediction Label
highlight its superior capability in handling high-dimensional
Figure 4: Confusion Matrix for Custom CNN data and complex histopathological features, outperforming
most deep learning models in this study. Its ability to balance
Figure 4 displays the confusion matrix for the Custom precision and recall while maintaining a low error rate
CNN model. The higher number of false positives and false emphasizes its robustness as a classification tool in medical
negatives compared to the Random Forest model indicates diagnostics. Despite being a traditional machine learning
model, Random Forest's performance showcases its strength
that Custom CNN requires further optimization to improve
when coupled with strong feature extraction techniques. The
its diagnostic accuracy. Enhancements in model architecture
custom CNN, although yielding a lower accuracy of 79.99%,
and training data may be necessary to boost performance. provided valuable insights into the potential of convolutional
architectures for feature extraction and image classification.
The confusion matrix for ResNet50, in Figure 5, shows a
As a custom-built model, it highlighted the effectiveness of
relatively balanced distribution of true positives and true CNNs in learning spatial hierarchies from image data,
negatives, with relatively few false positives and false particularly in medical imaging tasks. Its performance
negatives. suggests that while there is room for improvement, CNN-
Confusion Matrix based models remain a promising approach, especially when
further optimized with larger datasets or deeper architectures.
Benign 428 172 Pre-trained deep learning models, such as VGG-16 and
True Label

ResNet-50, also demonstrated competitive performance, with

Malignant 173 891 accuracy of 86.96% and 95.95%, respectively. These models
benefitted from transfer learning, leveraging their pre-trained
Benign Malignant
knowledge on extensive datasets like ImageNet to improve
feature extraction in the breast cancer detection task.
Prediction Label
Future research should address several key challenges in
Figure 5: Confusion Matrix for ResNet50 improving breast cancer detection. One promising direction
involves exploring multi-modal data integration, combining
This indicates good overall performance, with room for histopathological images with other types of medical data
improvement in reducing errors. such as genetic information, patient history, or radiological
Confusion Matrix images. This integration could enhance diagnostic accuracy
and lead to more personalized treatment plans, pushing the
Benign 493 107 boundaries of precision medicine. Additionally, Federated
Learning could be employed to facilitate collaborative model
True Label

Malignant 140 924 development across different medical institutions while

preserving patient data privacy. This decentralized learning
Benign Malignant approach allows models to learn from diverse datasets without
the need for data centralization, thus addressing concerns
Prediction Label related to data security and compliance with medical
regulations.
Figure 6: Confusion Matrix for VGG16
Furthermore, investigating advanced architectures like
Vision Transformers (ViTs) presents another exciting
The confusion matrix for VGG16 demonstrates its high
opportunity. ViTs, which apply transformer models to vision
precision, with a lower number of false positives. However,
tasks, have shown potential in surpassing CNNs in some
there are some false negatives, indicating that the model might
image classification benchmarks. Hybrid models that combine
miss some malignant cases.
CNNs with transformers may offer a new frontier in medical Learning,” 2022, pp. 121–127. doi: 10.1007/978-981-16-
image analysis by merging the strengths of both 2008-9_11.
architectures—CNNs’ local feature extraction and [5] M. H. Alshayeji and J. Al-Buloushi, “Breast Cancer
Classification Using Concatenated Triple Convolutional
transformers’ global attention mechanisms. These hybrid Neural Networks Model,” Big Data and Cognitive
approaches could lead to improved accuracy, robustness, and Computing, vol. 7, no. 3, Sep. 2023, doi:
generalizability in detecting complex patterns within medical 10.3390/bdcc7030142.
images. [6] M. Humayun, M. I. Khalil, S. N. Almuayqil, and N. Z.
Jhanjhi, “Framework for Detecting Breast Cancer Risk
By pursuing these advanced techniques and Presence Using Deep Learning,” Electronics (Switzerland),
methodologies, the field of medical imaging could see vol. 12, no. 2, Jan. 2023, doi: 10.3390/electronics12020403.
significant advancements, ultimately leading to the [7] P. Pratheep Kumar, V. Mary Amala Bai, and G. G. Nair,
development of more accurate, reliable, and secure diagnostic “Augmentation Techniques on Mammogram Images for
tools. Such innovations have the potential to revolutionize CNN Based Breast Cancer Classification,” in Lecture Notes
in Electrical Engineering, Springer Science and Business
breast cancer detection, offering better outcomes for patients Media Deutschland GmbH, 2022, pp. 635–649. doi:
through early diagnosis, tailored treatment plans, and more 10.1007/978-981-16-8774-7_53.
precise medical interventions, thus propelling the future of [8] A. S. Elkorany and Z. F. Elsharkawy, “Efficient breast cancer
patient-centered care in oncology. mammograms diagnosis using three deep neural networks
and term variance,” Sci Rep, vol. 13, no. 1, Dec. 2023, doi:
REFERENCES 10.1038/s41598-023-29875-4.
[1] A. Oliver, X. Lladó, J. Freixenet, and J. Martí, “False [9] T. M. Mengistu, B. S. Arega, and B. H. Belay, “Transfer
Positive Reduction in Mammographic Mass Detection Using Learning with Pre-trained CNNs for Breast Cancer Stage
Local Binary Patterns,” in Medical Image Computing and Identification,” in Lecture Notes of the Institute for
Computer-Assisted Intervention – MICCAI 2007, Berlin, Computer Sciences, Social-Informatics and
Heidelberg: Springer Berlin Heidelberg, pp. 286–293. doi: Telecommunications Engineering, LNICST, Springer
10.1007/978-3-540-75757-3_35. Science and Business Media Deutschland GmbH, 2023, pp.
127–136. doi: 10.1007/978-3-031-28725-1_8.
[2] N. Tirada et al., “Breast Cancer Tissue Markers, Genomic
Profiling, and Other Prognostic Factors: A Primer for [10] D. Kumar and U. Batra, “Breast Cancer Histopathology
Radiologists,” RadioGraphics, vol. 38, no. 7, pp. 1902–1920, Image Classification Using Soft Voting Classifier,” in
Nov. 2018, doi: 10.1148/rg.2018180047. Lecture Notes in Networks and Systems, Springer Science
and Business Media Deutschland GmbH, 2021, pp. 619–631.
[3] O. Jimenez-del-Toro et al., “Analysis of Histopathology doi: 10.1007/978-981-15-9712-1_53.
Images,” in Biomedical Texture Analysis, Elsevier, 2017, pp.
281–314. doi: 10.1016/B978-0-12-812133-7.00010-7.
[4] A. Pillai, A. Nizam, M. Joshee, A. Pinto, and S. Chavan,
“Breast Cancer Detection in Mammograms Using Deep