Source-linked AI summary
Breast cancer detection using artificial intelligence techniques: A systematic literature review
Ali Bou Nassif, Manar Abu Talib, Qassim Nasir, Yaman Afadar, Omar Elgendy
TL;DR
Breast-cancer detection research spans genetic and imaging data, but radio-genomics is rarely used because its datasets are expensive and computationally demanding. This systematic review examined 80 machine-learning and deep-learning studies and found CNNs widely used with generally good results.
Problem
Breast-cancer detection research spans genetic and imaging data, while radio-genomics remains rarely used because its datasets are expensive and computationally demanding.
Method
The paper systematically reviews machine-learning and deep-learning studies comparing genetic sequencing and imaging data, models, and feature-extraction methods for breast-cancer detection.
Results
The review identified 80 relevant papers and found CNNs widely used for gene-expression and MRI data, often obtaining good results versus other algorithms.
Takeaways & Limitations
Future studies should explore hybrid and underinvestigated deep-learning models, merge gene-sequencing datasets, and use feature selection for broader outcomes.
Takeaways & Limitations
Most reviewed studies evaluated performance mainly with accuracy, often omitting confusion-matrix parameters and AUC, which the authors consider insufficient.
Abstract
from arXiv · showhide
Cancer is one of the most dangerous diseases to humans, and yet no permanent cure has been developed for it. Breast cancer is one of the most common cancer types. According to the National Breast Cancer foundation, in 2020 alone, more than 276,000 new cases of invasive breast cancer and more than 48,000 non-invasive cases were diagnosed in the US. To put these figures in perspective, 64% of these cases are diagnosed early in the disease's cycle, giving patients a 99% chance of survival. Artificial intelligence and machine learning have been used effectively in detection and treatment of several dangerous diseases, helping in early diagnosis and treatment, and thus increasing the patient's chance of survival. Deep learning has been designed to analyze the most important features affecting detection and treatment of serious diseases. For example, breast cancer can be detected using genes or histopathological imaging. Analysis at the genetic level is very expensive, so histopathological imaging is the most common approach used to detect breast cancer. In this research work, we systematically reviewed previous work done on detection and treatment of breast cancer using genetic sequencing or histopathological imaging with the help of deep learning and machine learning. We also provide recommendations to researchers who will work in this field
1. Introduction
The introduction presents breast cancer as a major cause of death in women and outlines its main disease types. It also motivates radio-genomics as an approach linking imaging and genetic information to support diagnosis, prediction, and early detection.
- Background: Breast cancer is a major cause of death among women worldwide and occurs in four main types: normal, benign, in-situ carcinoma, and invasive carcinoma.The passage reports that 41,760 women and more than 500 men recently died from breast cancer.
- Motivation: Radio-genomics is an emerging field focused on multi-scale associations between medical imaging and gene expression data.It addresses the relationship between tissue-level imaging and underlying molecular and genetic disease biomarkers.
- Motivation: Radio-genomics provides radiological and genetic features that can enhance diagnosis, analyze tissues at the molecular level, and support cancer prediction and early detection.The passage contrasts this with less precise imaging, which may lead to over- or under-treatment.
1. Which deep learning models perform most effectively?
This section frames the review around comparing deep learning with classical machine learning models and identifying the metrics used to assess performance. It also covers features, datasets, and the trade-offs between gene sequencing and imaging for breast cancer detection.
- Model performance: The review compares deep learning and classical machine learning models using reported performance metrics.It aims to assess which deep learning models perform most effectively.
- Features: The review identifies important features for breast cancer classification and examines how those features are selected and extracted.Feature selection and extraction methods are explicitly considered.
- Datasets: The review catalogs public and private datasets for gene sequencing and MRI imaging data.It also lists methods used to select and extract features from these datasets.
- Data modalities: The review compares gene sequence data with image data for breast cancer detection, emphasizing their drawbacks, challenges, and advantages.The differences between the two approaches are presented in tabular form.
2. Related Work
Prior studies examined breast cancer detection through imaging or genomics, but none, according to the authors, included both techniques. This work addresses that gap by studying genetic sequencing and imaging together and offering recommendations for future research.
- No prior research had included both imaging and genomics for breast cancer detection, according to the authors.
- Histopathological imaging studies found ANNs and PNNs were frequently applied, while textural and morphological features dominated feature extraction.Deep CNNs were reported as effective for early detection and diagnosis.
- Prior reviews reported imaging was used in most breast cancer studies, whereas genetics appeared in only a few.The reviewed literature included 27 machine-learning papers, 4 ensemble-technique papers, and 8 deep-learning papers.
- Genetic breast cancer detection research used gene annotation, gene finding, and mutation detection with regression, probability models, SVMs, NNs, and deep learning.
- Deep-learning reviews covered ultrasound, mammography, and MRI, organizing studies by datasets, architectures, applications, and evaluation.These studies examined deep-learning-based computer-aided diagnosis systems for breast imaging.
- The authors’ contribution combines genetic sequencing and imaging to predict breast cancer, support early diagnosis and treatment, and provide research recommendations.
3. Methodology
The review systematically examined recent artificial-intelligence research on breast cancer detection, focusing on deep learning, machine learning, genetic data, and medical imaging. It searched major academic databases, screened more than 1,000 publications, and selected approximately 80 studies for structured analysis.
- Scope: Approximately 80 recent papers on breast cancer diagnosis and treatment were reviewed, covering deep learning alone and hybrid machine-learning/deep-learning approaches.The review targeted breast cancer detection using deep learning and included papers using either deep learning or combined machine learning and deep learning.
- Search strategy: The Scopus database was used primarily to obtain refereed articles, with selected papers distributed across PubMed, ScienceDirect, IEEE, Springer, and Nature.The paper states that Scopus was the main database and identifies the five leading databases in the selected-paper distribution.
- Search strategy: More than 1,000 papers published between January 2010 and May 2020 were identified using a search combining breast cancer, deep learning, genetic, and imaging terms.The search statement included genes, genomes, microarrays, DNA, X-ray, mammography, MRI, and ultrasound.
- Eligibility criteria: The screening reduced the corpus to 80 journal and conference papers focused on genetic expression, gene sequencing, and ultrasound, radiography, mammography, or MRI.Breast cancer study publications peaked around 2019 and 2020 before the selection was narrowed.
- Data extraction: The review extracted publication details, algorithms, learning approach, accuracy and other performance measures, datasets, and features to answer its research questions.The recorded information included whether studies used deep learning alone or a hybrid deep-learning/machine-learning approach.
4. Results and Discussion
The systematic review screened 1,000 conference and journal papers, ultimately retaining 80 studies related to machine learning and deep learning, with selection focused on deep learning approaches or hybrid DL-ML models.
- Study selection: 1,000 conference and journal papers were identified initially, but duplicate, unrelated, and purely medical or general-cancer studies were removed.The screening process reduced the initial search to studies relevant to breast-cancer detection using ML and DL.
- Study selection: 80 papers related to both machine learning and deep learning remained after screening.The retained literature formed the evidence base for the systematic study.
- Study selection: The review specifically selected papers addressing deep learning approaches or deep-learning–machine-learning hybrid models.Figure 3 describes the information flow through the systematic-review phases.
1. Which deep learning models perform most effectively?
CNN is the most frequently used model and performs strongly across gene-expression and imaging data, while binary classification generally achieves higher accuracy than multiclass classification. Standalone deep-learning models outperform hybrid models in gene-expression studies, whereas imaging-based subtype classification reaches 90% accuracy at its highest.
- Binary classification generally achieves higher accuracy than multiclass classification in breast cancer detection studies.
- CNN is the most frequently used model for both binary and multiclass breast cancer classification.
- 90% was the highest accuracy reported for multiclass differentiation or breast cancer subtype classification using imaging data.
- CNN delivers excellent performance for both gene-expression and imaging data, including 97% accuracy for binary classification.
- Standalone deep-learning models consistently achieve higher accuracy than hybrid models on gene-expression data.The hybrid model in paper achieved 87% accuracy, while CNN in paper achieved 95% and BPNN achieved 99.8%.
- SOCPNN and CNN are identified as the models with the highest accuracies in Figures 5 and 6, respectively.
2. What datasets are available for gene sequencing and MRI?
The review identified public and private datasets for gene sequencing and breast-cancer imaging, with more imaging datasets available than genetic datasets. The Cancer Genome Atlas was the most used gene-expression dataset, while Wisconsin and MRI datasets were among the main imaging resources used.
- Gene-sequencing datasets: The Cancer Genome Atlas was the most used gene-expression dataset and includes data from many cancer types, including breast cancer.It aims to identify DNA changes across cancers and includes clinical information for each participant.
- Dataset availability: Public and private datasets were available for both gene sequencing and breast-cancer imaging, but imaging datasets were more numerous.Many imaging datasets were public and free, whereas few publicly available gene-sequencing datasets were not free.
- Imaging datasets: The Wisconsin breast cancer dataset was the most used imaging dataset and contains features computed from digitized fine-needle-aspirate images of breast masses.The dataset was obtained from the UCI repository.
- Imaging datasets: Large imaging datasets such as DDSM can be used independently, whereas smaller datasets can be merged with another dataset to improve the data pool.Most other breast-cancer imaging datasets were public and free.
- Imaging datasets: Researchers mainly used the public Wisconsin breast cancer and MRI datasets because both contained many samples.The review also listed public imaging resources including DDSM, MIAS, and other datasets.
3A. What are the most commonly used features for breast cancer classification?
Breast cancer classification commonly uses tumor information, protein status, and imaging-derived morphological features. These features describe tumor characteristics, cancer-related proteins, lymph-node involvement, metastasis, and cellular structure.
- Tumor features include tumor size and grade, with grade depending on cancer type.
- Protein and disease-status features include ER status, PR status, positive lymph-node count, metastasis sites, and HER2 status.Positive lymph nodes record how many nodes contain cancer, while metastasis records the number of sites to which cancer has spread.
- Imaging classification uses general morphological features such as symmetry, compactness, and concave points.
- Breast-image features include marginal adhesion, uniformity of cell size, and clump thickness.Uniformity of cell size is represented by 0 when lacking uniformity and 1 when uniform; clump thickness describes mono- or multi-layered cells.
3B. What are the most effective feature selection and feature extraction methods?
CNNs and other feature-selection and extraction methods support breast-cancer imaging and genetic-data analysis, but imaging is easier to process while genetic data may be more accurate. Each datatype nevertheless has distinct drawbacks and challenges.
- Feature extraction methods: Feature-Based Strategy identifies localized image features, including edges and corners, and tracks their movement across consecutive frames.The strategy involves finding features in two or more consecutive images before tracking them.
- Feature extraction methods: CNNs support feature extraction through convolutional layers and classification through fully connected layers.Feature extraction commonly uses either the last hidden layer or the flattened last convolutional layer.
- Feature selection methods: XGBoost and random forests were used for selecting features from mRNA and protein expression profiles.Random forests rank features according to how well they improve node purity.
- Data-type comparison: Imaging data offers more easily extracted features and more preprocessing options, but CNNs may extract nonrelated features that require additional cleaning.These unrelated features can contribute to poor performance.
- Data-type comparison: Genetic expression data contains fewer features but may be more effective and is consistently more accurate, especially for multiclass prediction.The best imaging-data performance reported was 99.3% accuracy.
- Data-type comparison: Genetic data processing is complex and expensive, and sufficiently large, correctly labeled datasets are difficult to find.The survey concludes that neither datatype can be declared universally superior because both have drawbacks and challenges.
5. Conclusions and Future Research Directions
The review identifies underused deep learning mechanisms and evaluation measures in breast cancer detection and subtype classification. It recommends hybrid CNN and Attention approaches, broader genetic datasets and outcomes, and expanded multiclass research.
- Research gaps: Most papers use machine learning, while deep learning remains underinvestigated, creating opportunities to apply LSTM, GAN, and RNN models.These mechanisms have not yet been used in this field, according to the review.
- Evaluation: Most studies emphasize Accuracy while ignoring confusion matrix parameters and AUC, so future work should include AUC and F-scores.Accuracy does not distinguish false positive from false negative classifications.
- Model development: CNN is widely used for gene expression and MRI images and often obtains good results, motivating further development of hybrid CNN algorithms.The review identifies CNN as an important finding across both data types.
- Model development: Attention has been used infrequently for image classification, providing an opportunity to improve deep learning model accuracy with this mechanism.The passage specifically presents Attention as a direction for future researchers.
- Genetic data: Future studies can merge multiple gene sequencing datasets, predict additional outcomes with larger datasets, and use feature extraction and selection to improve performance.Confusion matrix parameters are also proposed to increase accuracy, while feature selection can eliminate non-significant features.
- Future applications: Research should extend beyond detection and subtype classification to risk levels, recurrence prediction, and multiclass predictors using genetic data.Most reviewed genetic-sequencing studies used binary classification, primarily for breast cancer detection and likelihood.