Image analysis with deep learning to predict breast cancer grade, ER status, histologic subtype, and

intrinsic subtype

This study examines the use of deep learning to analyze H&E-stained histologic images to predict critical
breast cancer markers, including tumor grade, ER status, histologic subtype, and intrinsic subtype. A
model trained on 571 tumors showed high accuracy in classifying tumor grade (82%), ER status (84%),
and other key markers. The approach demonstrated promise for the identification of patients requiring
genomic testing or replacing traditional scoring methods, especially in resource-limited settings where
RNA-based tests are not accessible. ER status misclassification was more common in Luminal B tumors,
indicating a potential area of improvement. This method presents a cost-effective alternative for clinical
decision-making.

Multi-Class Breast Cancer Classification using Deep Learning Convolutional Neural Network

This work presents a deep learning approach to multi-class classification of breast cancer using a
DenseNet CNN model. It predicts the specific tumor subtypes, such as Fibroadenoma and Lobular
carcinoma, rather than benign or malignant categories, which is not predicted by traditional methods.
The experiments conducted on the BreakHis dataset demonstrated an accuracy of 95.4%, surpassing
state-of-the-art models. This work thus points to the potential of automatic image processing
techniques in enhancing the diagnosis of breast cancer.

A deep learning image-based intrinsic molecular subtype classifier of breast tumors reveals tumor
heterogeneity that may affect survival

This study establishes an approach using deep learning for approximating intrinsic PAM50 subtyping of
breast cancer through the use of images from biopsy tissues stained by H&E. Training was carried out on
443 tumors while validation was carried out on 222 samples where there was precise identification of
major molecular subtypes with Basal-like and Luminal A classification but with major intratumoral
heterogeneity in comparison between the two subtypes. Such subtypes showed heterogeneous survival
rates as well as diverse hormone receptor expressions. The proposed method reduces human effort in
determining cancer-rich patches and indicates deep learning can potentially approximate RNA-seq-
based molecular tests while achieving better tumor heterogeneity detection.

Classifying Breast Cancer Subtypes Using Deep Neural Networks Based on Multi-Omics Data

It brings forth a novel deep learning model, DeepMO, that pools together multi-omics data-including
mRNA, DNA methylation, and copy number variation-and uses them in the classification of breast cancer
subtypes. By preprocessing and selecting features of these data types, DeepMO improved over other
approaches in terms of binary classification accuracy and AUC. Besides, it exhibits superior performance
for multi-classification than methods involving single or multi-omics data. It also investigated how
feature selection could impact the outcomes and identified several genes associated with biologically
important pathways. The findings indicate that DeepMO is a great tool for studying multi-omics data in
the context of breast cancer research.
Research Article RNA-Seq-Based Breast Cancer Subtypes Classification Using Machine Learning
Approaches

This research has focused on the complexity of BRCA, with several subtypes that differ through their
interaction mechanisms at the gene level. In this research, the authors develop a method using
weighted DEGs and gene co-expression networks for identifying biological significance and interaction
mechanisms for each subtype. The performance of their weighted DEGs in the binary classification
model for predicting unseen samples is very strong. In addition, they developed a new Gene Ontology
enrichment method, GOEGCN, which provides insight into the biological functions affected by
modulated gene interactions. The results indicate that gene networks play a crucial role in
understanding BRCA subtypes and improving personalized treatment approaches.

A Cascade Deep Forest Model for Breast Cancer Subtype Classification Using Multi-Omics Data

This paper discusses the application of the cascade Deep Forest ensemble model to classify breast
cancer subtypes: IntClust and Pam50, using multi-omics data. Unlike many other methods which
integrate multiple sources of data with a cost on computational performance, this approach centers on
gene expression data, resulting in high accuracy: 83.45% for 5 subtypes and 77.55% for 10 subtypes. The
model outperformed other techniques, especially for imbalanced datasets, with significantly fewer
computational times; it used 5 seconds to run 10 subtypes and 7 seconds on 5 subtypes. The results,
therefore, shine light on the prospects of cascade Deep Forest for efficient and accurate classification of
breast cancer subtype.

A hybrid gene selection model for molecular breast cancer classification using a deep neural network

This study introduces a deep neural network-based model that is applied in the molecular classification
of breast cancer based on the use of microarray gene expression data. A hybrid approach between filter
and wrapper-based feature selection was incorporated to ensure that relevant genes for classification
purposes were selected. In using six distinct microarray datasets, the model demonstrated excellent
performance in achieving accuracy with sensitivity, specificity, f-measure, MCC, and recall surpassing
other approaches. The results confirmed that deep neural networks are very effective for breast cancer
molecular subtype classification and produce strong results across all the datasets selected.

Combination Ultrasound and Mammography for Breast Cancer Classification using Deep Learning

This paper suggests a hybrid approach for breast cancer classification by using deep learning to enhance
the accuracy of ultrasound and mammography. The technique consists of two phases: the first is the
random combination of images through the k-combination technique, and the second is the
classification of images using MobileNet deep learning. The outcome shows that the combination
technique increases the accuracy rate by reducing false positives by 13% and false negatives by 14%.
This will improve the possibility of avoiding unnecessary biopsies and surgeries while making treatment
planning more accurate.
Deep learning with convolutional neural network in the assessment of breast cancer molecular
subtypes based on US images: a multicenter retrospective study

This work assesses the efficacy of a deep convolutional neural network (DCNN) for breast cancer
subtyping based on ultrasound (US) images. A total of 4828 US images of 1275 patients were used in the
training process and were successful in identifying four different molecular subtypes, with
differentiation of luminal from non-luminal disease. The model had high accuracy in two independent
test cohorts, with accuracy ranging from 80% to 97%, and strong performance in classifying luminal
disease. Overall, the DCNN model showed excellent potential for assessing molecular subtypes of breast
cancer based on pre-treatment US images.

Evaluating the Accuracy of Breast Cancer and Molecular Subtype Diagnosis by Ultrasound Image Deep
Learning Model

This research tests an optimized DLM to predict the molecular subtype of breast cancer based on
ultrasound images. There were 3,739 images that came from two hospitals, for which training and
testing were carried out on the model. DLM has proved superior over traditional BI-RADS, in which
better accuracy, at 89.7%, with better sensitivity of 91.3% and specificity of 86.9% were established for
molecular subtype identification. The model showed strong performance in predicting subtypes such as
triple-negative and HER2 (+) with AUC values ranging from 0.81 to 0.86. It also demonstrated potential
to reduce unnecessary biopsies by 67.86%, though the false negative rate was 10.4%.

Improved breast cancer histological grading using deep learning

This paper presents DeepGrade, a deep learning model that enhances the risk stratification of patients
with Nottingham Histological Grade (NHG) 2 breast cancer. The model classifies NHG 2 into two groups:
DG2-high and DG2-low. Using digital whole-slide histopathology images from 1567 patients, the model
demonstrated that DG2-high patients have a significantly higher risk of recurrence in both internal and
external test sets. This new model provides more prognostic value than traditional histological grading,
offering more precise clinical information and a cost-effective alternative to molecular profiling for
treatment decisions.

Machine learning integrated ensemble of feature selection methods followed by survival analysis for
predicting breast cancer subtype specific miRNA biomarkers

This study investigates the role of microRNAs (miRNAs) in breast cancer, identifying significant miRNA
biomarkers across different subtypes using Next Generation Sequencing (NGS) data. The approach used
a two-phase strategy, combining machine learning for feature selection and survival analysis to identify
miRNAs linked to breast cancer survival. The top-performing miRNAs were categorized, and the study
found 27 miRNAs that were most significant for survival. Further analysis, including regulatory networks
and protein-drug interaction mapping, suggests potential therapeutic targets, providing insights for
future clinical trials in breast cancer treatment.
Multi- class classification of breast cancer abnormalities using Deep Convolutional Neural Network
(CNN)

This study addresses the challenge of early detection of breast cancer, which is very important in
reducing mortality. It applies the technique of developing a Deep Convolutional Neural Network that can
primarily distinguish, classify, and segment calcifications, masses, and carcinomas by using mammogram
images for diagnosis. Compared to prior methods, it moves towards more detailed detection of
malignancies instead of just deciding whether to classify them as benign or malignant. This proposed
model uses transfer learning with ResNet50 and changes the learning rate during training. The model
managed to achieve an accuracy rate of 88% for the classification of these abnormalities, thus providing
improved disease management.

Random Forest Modelling of High-Dimensional Mixed-Type Data for Breast Cancer Classification

This study develops a computational framework designed to integrate various types of data for the
analysis of complex diseases like breast cancer. By applying this framework to breast cancer data, the
researchers identified six distinct subtypes of triple-negative breast cancer with unique molecular
mechanisms. The framework also highlighted a group of patients who responded poorly to standard
chemotherapy, offering potential for better treatment strategies. The approach showed strong
performance and reproducibility when applied to two independent breast cancer cohorts,
outperforming traditional classification methods in predicting clinical outcomes.

The impact of scan number and its preprocessing in micro-FTIR imaging when applying machine
learning for breast cancer subtypes classification

This study applies FTIR imaging combined with machine learning to enhance the classification of breast
cancer molecular subtypes, such as Luminal B and HER2, with the hope of reducing subjectivity in
traditional immunohistochemistry. The results show that with variation in scan numbers and
preprocessing techniques, the quality of the spectrum improves, and variability reduces with an increase
in the scan number. Among six machine learning models tested, Extreme Gradient Boosting achieved
the highest accuracy with 0.995, with the b256_064 scan setting identified as optimal due to its balance
between accuracy and minimal acquisition time. This approach shows promise for more reliable cancer
classification and could serve as a foundation for future cancer research using machine learning.

XOmiVAE: an interpretable deep learning model for cancer classification using high-dimensional omics
data
XOmiVAE is an innovative deep learning model designed to address the lack of explainability in cancer
classification using high-dimensional omics data. By using a variational autoencoder (VAE), it provides
clear insights into the contribution of each gene and latent dimension in classification predictions as well
as their correlations. The model also explains unsupervised clustering results, thus making it one of the
first to interpret novel clusters generated by VAE. Validation through downstream tasks and existing
biomedical knowledge confirms that XOmiVAE's explanations align with current biological
understanding, offering a promising approach for new insights in cancer research.
A Hybrid Deep Learning Model for Predicting Molecular Subtypes of Human Breast Cancer Using
Multimodal Data

This study introduces a hybrid DL approach for predicting the subtypes of breast cancer from gene and
image modality data. The multimodal fusion model integrates information relating to patient genetics
and imaging data to enhance prediction accuracy. In the TCGA-BRCA dataset, a feature extraction
network is used for the different types of data and then weighted linear aggregation was applied to fuse
the output results. With a PCA to reduce gene data dimensions and improved data filtering for images,
the model achieved a prediction accuracy of 88.07%, which is more effective than traditional DL models.

Breast Cancer Mammograms Classification Using Deep Neural Network and Entropy-Controlled Whale
Optimization Algorithm

This paper investigates enhanced techniques for detecting and classifying breast cancer in CAD systems
to achieve higher accuracy, fewer false positives. This method applies the optimization of deep features
using a fine-tuned model like MobilenetV2 and Nasnet Mobile with the Modified Entropy Whale
Optimization Algorithm. By applying this technique to datasets such as INbreast, MIAS, and CBIS-DDSM,
the method achieved impressive accuracies of 99.7%, 99.8%, and 93.8%, respectively. The results were
compared with existing approaches, showing the superior performance of the proposed system in
breast cancer classification.

Classification of breast cancer subtypes: A study based on representative genes

This study describes the classification of breast cancer subtypes by adopting various machine learning
techniques and with the use of the PAM50 gene list. The support vector machine, it was observed by the
study using a linear kernel, had given the highest value of F1 scores as that of 0.98 achieved for the Basal
subtype with a poor prognostic one and 0.90 Her2 subtype being a worse type. Moreover, the research
includes gene analysis using SHAP values, which are used to identify the most important genes for the
classification of each subtype, thus leading to far more accurate and personalized breast cancer
treatments.

Deep learning approach for cancer subtype classifcation using high-dimensional gene expression data

This paper introduces a deep learning model, DCGN, that is designed to enhance the classification of
cancer subtypes. DCGN combines a convolutional neural network (CNN) and bidirectional gated
recurrent unit (BiGRU) to reduce the dimensionality of gene expression data and extract meaningful
features. It addresses the problem of small sample sizes and sparse data by applying synthetic minority
oversampling and handling high-dimensional features effectively. DCGN significantly excels other
methods when it comes to the classification of subtypes in breast and bladder cancers, including
accuracy and reliability better than the existing seven approaches.
Investigating Deep Learning Based Breast Cancer Subtyping Using Pan-Cancer and MultiOmic Data

The paper addresses the potential of using deep learning for breast cancer subtyping based on multi-
omic data, including microRNA expressions and copy number alterations. The authors compare
supervised and semi-supervised models, finding that simpler models performed just as well as deep
learning approaches using gene expression data. Deep learning models showed only minor
improvements in accuracy when combining multi-omic data, indicating the need for much larger
datasets for more accurate results. Linear models confirm known gene-subtype relationships, but deep
learning models are able to find non-linear relationships, which offers new insights into breast cancer
subtyping.

Performance Analysis of XGBoost Ensemble Methods for Survivability with the Classification of Breast
Cancer

This research focuses on the improvement of early breast cancer detection using machine learning
techniques. In the first phase, to handle the problem of imbalanced data, the Synthetic Minority
Oversampling Technique (SMOTE) is applied, and then classification is performed using Naive Bayes,
decision trees, Random Forest, and their ensembles. The best accuracy obtained was 98.20% by the
XGBoost-Random Forest ensemble. According to the study, ensemble methods as a machine learning
approach are extremely efficient in recognizing breast cancer very early.

Predicting the molecular subtype of breast cancer and identifying interpretable imaging features using
machine learning algorithm

This paper presents the application of interpretable machine learning models in predicting breast cancer
molecular subtypes from clinical and imaging data from 600 patients. Among the tested models, the
decision tree model performed best with high accuracy in distinguishing triple-negative breast cancer at
94.7% and sensitivity at 90.5%. The study also proved that the DT model enhanced the performance of
radiologists in the diagnosis of TNBC and Luminal subtypes, especially for less experienced radiologists.
The SHAP technique was applied to identify the key features that enhance the interpretability of the
model. This approach provides precious assistance to radiologists in breast cancer diagnosis.

Raman spectroscopy and machine learning for the classification of breast cancers

This paper presents the application of Raman spectroscopy coupled with machine learning methods to
classify subtypes of breast cancer and distinguish cancerous cells from normal cells. Two machine
learning algorithms, PCA-DFA and PCA-SVM, were applied to Raman spectra obtained from cultured
breast cancer cell lines. Both algorithms differentiated normal and cancer cells with over 97% accuracy
and achieved more than 92% accuracy in classifying subtypes of breast cancer. The findings emphasize
the possibility of Raman spectroscopy and machine learning in the noninvasive, accurate diagnosis of
breast cancer and classification of its subtypes.
Discriminative Dictionary Learning Using Penalized Rank-1 Approximation for Breast Cancer
Classification With Imbalanced Dataset

This paper introduces a new approach called Penalized Sequential Discriminative Dictionary Learning
(PSDDL) to enhance feature extraction for the classification of histopathological images, especially in the
context of breast cancer diagnosis. The proposed approach is based on dictionary learning and sparse
coding techniques to construct class-specific dictionaries that incorporate label information to improve
the accuracy of classification. To handle class imbalance, under-sampling is applied to the dataset.
Experimental results demonstrate that PSDDL outperforms existing methods for classifying the BreakHis
breast cancer dataset. Class-specific information enriches the interpretability of machine learning
models and aids in understanding complex biological tissue structures.

HER2 classification in breast cancer cells: A new explainable machine learning application for
immunohistochemistry

This paper discusses the application of explainable machine learning to improve the classification of
HER2 photomicrographs in breast cancer diagnosis. HER2 evaluation by IHC is qualitative and has
variability, thus requiring additional techniques such as FISH. A logistic regression-based ML model was
trained on 393 IHC microscopy images from 131 patients, and the model was able to distinguish
between upregulated and normal HER2 expression. This study shows that integrating ML improves the
diagnostic accuracy of results to support better treatment decisions when compared with IHC alone to
the combined IHC + FISH diagnosis.

Novel Deep CNN Model based Breast Cancer Classification

This paper introduces a novel DeepCNN model for the classification of breast cancer, which should
improve the precision in the diagnosis of the disease. Using deep learning techniques, the performance
of the model was optimized through hyper-parameter tuning using Random Search and factors such as
learning rate, dropout rate, and number of epochs. The results show that the custom DeepCNN model
outperforms several pre-trained models, such as VGG19, Resnet50, and Xception, with a magnificent
accuracy of 99.18%. The study shows the possibility of deep learning in raising the performance level of
diagnosis in breast cancer.

Patient Graph Deep Learning to Predict Breast Cancer Molecular Subtype

This study focuses on predicting breast cancer molecular subtypes with deep graph learning that
incorporates a multiplicity of diagnostic data sources. The methodology transforms patient data into a
multi-relational graph and extracts the features from DCE-MRI radiographic images and genomic
variants through autoencoders. Subtype prediction with this model leverages multimodal diagnostic
information within a Relational Graph Convolutional Network architecture. Results indicate that addition
of heterotypic data types improves the model's ability to classify subtypes and clearly shows the
potential of graph neural networks in multimodal data fusion in the analysis of breast cancer.
A Deep-Learning-based Novel Method to Classify Breast Cancer

This study focuses on improving the detection of breast cancer by combining DNN and ML algorithms,
with an impressive accuracy of 98.25%. The approach is a significant advancement in early breast cancer
detection by using data visualization, analysis, and pre-processing techniques. The proposed method can
revolutionize current diagnostic practices and improve patient outcomes and care. This study captures
the revolutionary power of advanced data analytics in the global battle against breast cancer.

A Multi-Task Transformer With Local-Global Feature Interaction and Multiple Tumoral Region
Guidance for Breast Cancer Diagnosis

This work proposes an advanced multi-task deep learning network that uses ultrasound images for
diagnosis of breast cancer. This combines a dual-stream encoder to integrate interactions between local
and global features using CNN as well as Transformer architectures, thereby improving both tumor
segmentation as well as classification. In addition, it includes a multi-tumoral region guidance module
that also helps in capturing long-range dependencies within tumor and its surroundings. By harnessing
both local and global features' strengths, this enhances the accuracy and efficiency of early-stage breast
cancer detection while minimizing possibly the time and professional expertise needed in diagnosing the
disease.

A Multi-View Deep Evidential Learning Approach for Mammogram Density Classification

This paper presents MV-DEFEAT, an evidential deep learning framework designed to enhance
mammogram density classification by emulating the decision-making process of a radiologist. It
integrates Dempster Shafer evidential theory with subjective logic to combine multiple mammogram
views for accuracy enhancement. The approach outperforms the state-of-the-art MVDL model,
achieving significant improvements in terms of area under the receiver operating curve (AUC) across
four open-source datasets. Apart from this, BIRADS classification and benign/malignant classification
show a fair amount of performance gain. MV-DEFEAT, based on the previous results, will be ready to be
clinically implemented as it depicts robust and reliable performance.

Automated Identification of Breast Cancer Type Using Novel Multipath Transfer Learning and
Ensemble of Classifier

In this research paper, the use of Artificial Intelligence and Machine Learning has been explored for
diagnosis in breast cancer by using deep learning along with traditional machine learning approaches.
The four regions of breast cancer images are divided, and the process followed by transfer learning with
ResNet50 and VGG16 was done for feature extraction. Then classification was done using various
models. Among these, the Voting Classifier resulted in the highest accuracy with 96.86%. The technique
classifies the various forms of breast cancer effectively, thus providing benefits such as early detection
and accurate diagnosis. The outcomes of this research demonstrate the ability of this method to
enhance breast cancer care and treatment strategies.
DEBCM: Deep Learning-Based Enhanced Breast Invasive Ductal Carcinoma Classification Model in
IoMT Healthcare Systems

This paper reports on early-stage breast cancer detection using a method based on convolutional neural
networks (CNN) for classifying histology images of invasive ductal carcinoma. The approach used
transfer learning and data augmentation to improve the predictive performance of the model. A fine-
tuning experiment with breast histology data and selection of the model via cross-validation brought an
excellent accuracy of 99.04%. The results outperform the baseline models in all evaluation metrics,
which makes it a strong candidate for early breast cancer detection in IoMT healthcare systems.

FastLeakyResNet-CIR: A Novel Deep Learning Framework for Breast Cancer Detection and
Classification

This paper proposes an improved ResNet architecture for the fast and effective detection and
classification of breast cancer called FastLeakyResNet-CIR. It managed to obtain 98.94% accuracy on the
BreakHis dataset, significantly better than all state-of-the-art approaches: ResNet18, ResNet50,
InceptionV3, and VGG16. This outcome presents a prospect of obtaining prompt and precise diagnosis
for patients affected by breast cancer.

Improved Butterfly Optimization Algorithm for Automated Breast Cancer Detection and Classification
using Deep Learning

This work introduces mammography images-based breast cancer diagnosis with the MIAS dataset. The
method is based on feature extraction and uses the Improved Butterfly Optimization Algorithm (IBOA)
for feature selection, along with deep adaptive spatial-based feature fusion. The proposed approach
obtained good performance with accuracy at 99.98%, sensitivity of 98.79%, specificity of 99.50%, and an
F1-score of 98.90%. This method outperformed prevalent techniques M-FD, IMPA-ResNet50, and IFSGA-
DNN for eliminating biases and enhancing the accuracy of the model.

StethoNet: Robust Breast Cancer Mammography Classification Framework

This paper proposes a deep learning framework for the detection of breast cancer using mammograms
called StethoNet. StethoNet was trained on CMMD and then evaluated on various independent
datasets, including Vindr-Mammo and INbreast. StethoNet attained remarkable AUC values of 90.7%,
83.9%, and 85.7%, respectively, over these three independent datasets. In order to deal with domain-
shift effects, entropy-based domain adaptation is applied along with Bayesian hyperparameter
optimization. The model was further validated through attention maps, which showed robustness and
generalizability in breast cancer detection.
SwinCNN: An Integrated Swin Transformer and CNN for Improved Breast Cancer Grade Classification
This paper presents a hybrid multi-class classification model for breast cancer detection based on
histopathological images. The model combines depth-wise separable convolutional networks and
transformers in order to account for both local and global features, thereby increasing the accuracy of
classification. The model was evaluated on three public datasets: BACH, BreakHis, and IDC, which
achieved impressive accuracies of 97.8%, 98.13%, and 98.32%, respectively. With this approach, it
addresses the self-attention complexity in transformers via patch merging that is robustly applicable to
the early diagnosis and grading of breast cancer.

USING SEGMENT-LEVEL ATTENTION TO GUIDE BREAST ULTRASOUND VIDEO CLASSIFICATION

This paper provides a new way to classify breast cancer from ultrasound videos without any human
keyframe selection by radiologists. The introduced method introduces the Segment Attention Generator
module that would allow deep learning models to focus their attention on video segments most relevant
for classification tasks. Results indicate that this indeed leads to huge improvements in performance in
classification tasks. In addition to this, it introduces a novel breast ultrasound video dataset and puts it
in the open access of not only the data but also its code.

convoHER2: A Deep Neural Network for MultiStage Classification of HER2 Breast Cancer

This study presents a computational model, convoHER2, designed to detect HER2 breast cancer using
convolutional neural networks (CNN) with hematoxylin and eosin (H&E) and immunohistochemical (IHC)
stained images. The model was trained and tested on a dataset of 4873 images and obtained accuracies
of 85% with H&E images and 88% with IHC images for the detection of HER2 cancer and its stages. The
result shows the possible use of the convoHER2 as an economic means of diagnosing HER2 breast
cancer, potentially improving patient care.