Breast cancer remains one of the leading causes of cancer-related death in women worldwide, with approximately 2.3 million diagnoses and nearly 29.8% mortality in 2020 alone. Despite progress in medical imaging and treatment, significant gaps remain in diagnostic accuracy and access to early detection, particularly in low-resource regions.
Artificial intelligence, and specifically deep learning, has emerged as a promising approach to improving breast cancer diagnosis. These techniques can detect subtle patterns in medical images such as mammograms, ultrasounds, and histopathology slides that may go undetected by human practitioners, enabling earlier and more consistent identification of malignant lesions.
This systematic review examines 62 peer-reviewed studies from IEEE, Scopus, and PubMed published between 2018 and 2023, evaluating the strengths, limitations, and accuracy of various deep learning models, with a particular focus on convolutional neural networks (CNNs) and their application across multiple imaging modalities in breast cancer detection.
Mammography is the primary screening tool for breast cancer, using X-ray technology to detect tumors before they become symptomatic. It is recommended for women aged 40 and above and has a sensitivity of approximately 67.8%, though effectiveness can vary by age, breast density, and operator skill.
Ultrasound is used as an adjunct to mammography, particularly in younger women and those with dense breasts. It achieves a specificity of 98% and is especially useful for distinguishing solid from cystic masses, though it is highly operator-dependent. MRI is reserved for high-risk individuals and provides sensitivity up to 96%, offering detailed views of dense tissue not visible by other means.
More advanced modalities including CT and PET scans are used for staging and monitoring treatment response, with CT achieving 91% sensitivity and 93% specificity. Image-guided biopsy, while more invasive, remains the gold standard for confirming diagnoses, providing detailed pathological staging information that guides treatment planning.
Convolutional Neural Networks (CNNs) are the most widely used deep learning architecture in breast cancer imaging. Their design principles of weight sharing, pooling, and local receptive fields make them highly effective at extracting hierarchical features from images. CNNs can classify lesions directly from raw medical images, but they require large labeled datasets to achieve high accuracy.
Several complementary architectures play important roles. Autoencoders compress and reconstruct image data, making them valuable for denoising mammograms, anomaly detection, and generating synthetic training images when labeled data is scarce. Generative Adversarial Networks (GANs) generate synthetic medical images that enrich training datasets and help improve model robustness, though they can be difficult to train due to instability and mode collapse.
Deep Belief Networks (DBNs) use unsupervised pre-training to learn representations from unlabeled data, which is valuable when annotated medical images are limited. Extreme Learning Machines (ELMs) offer computationally efficient classification with a single hidden layer, making them suitable for real-time applications. Each architecture offers distinct tradeoffs between accuracy, computational cost, and the amount of labeled data required.
This systematic review used the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) methodology to ensure transparency and reproducibility. The initial search identified 4,339 records from three databases, with the majority coming from Scopus (87.6%), followed by PubMed (10.9%) and IEEE (1.5%).
After removing 30 duplicates and screening out 3,945 irrelevant records, 364 articles were reviewed in detail. Studies were excluded based on publication year (older than five years), irrelevance to breast cancer, or focus on cancers other than breast cancer. This process narrowed the pool to 62 studies that met all inclusion criteria.
Data were extracted by independent reviewers, capturing author information, study design, classification method, accuracy metrics, and the dataset used. Research questions focused on identifying effective deep learning methods, evaluating their pros and cons, comparing performance metrics, and cataloging the most common breast cancer imaging datasets used in the literature.
Across the reviewed studies, reported accuracies ranged from 82.3% to 100%, demonstrating the wide capability of deep learning models for breast cancer detection. Several methods showed particularly strong performance: a DeepCNN model with random search optimization achieved 99.18% accuracy on multiple mammography datasets, and an AlexNet-based transfer learning model on MRI images reached 99.65%.
Transfer learning, where pre-trained models are fine-tuned on breast cancer datasets, consistently proved effective. By leveraging weights trained on large general image datasets like ImageNet, these models achieve high diagnostic precision even with limited domain-specific training data. This approach is particularly valuable in medical imaging, where large labeled datasets are expensive and time-consuming to produce.
Object detection frameworks like YOLOv3 and Faster-RCNN demonstrated exceptional precision in detecting and localizing lesions, achieving up to 99.17% precision on standard mammography datasets. These methods are especially valuable for real-time screening applications where both speed and accuracy are critical.
Across different imaging modalities, specific methods emerged as particularly effective. For digital mammography, DenseNet-II reached 94.55% accuracy, and a fuzzy U-Net segmentation approach on the DDSM dataset achieved a Jaccard index of 97.4% and 98.2% accuracy. For infrared thermography, an optimized LFR-COA-DenseNet121 model achieved near-perfect performance with 99.97% accuracy.
For histopathological images, a combined ensemble of PResNet-34, FE-VGG-16, and M-AlexNet with SVM classification achieved 99.99% accuracy, demonstrating that combining multiple feature extractors with classical classifiers can yield exceptional results. DCGAN-generated synthetic ultrasound images also enabled models to achieve 100% accuracy in benign vs. malignant classification, highlighting the value of data augmentation.
The most widely used benchmark datasets in breast cancer deep learning research are DDSM (Digital Database for Screening Mammography), used in 36.4% of studies, followed equally by INBreast and MIAS (Mammographic Image Analysis Society). Standardized datasets are essential for comparing model performance across studies, and their widespread adoption facilitates reproducibility.
The primary advantage of deep learning for breast cancer diagnosis is its high accuracy in image classification, with many CNN-based models exceeding 97% accuracy. These models can recognize complex patterns in mammograms and histopathology images that are difficult for human radiologists to detect consistently, reducing false positives and false negatives in screening programs.
A significant limitation is the dependency on large, well-labeled datasets, which are expensive and time-consuming to create in medical settings. Deep learning models trained on small or poorly balanced datasets are prone to overfitting, where the model memorizes training examples rather than learning generalizable patterns. Transfer learning partially addresses this limitation but does not eliminate it.
The black-box nature of deep neural networks remains a critical concern for clinical adoption. When clinicians cannot understand how a model reached its conclusion, trust and adoption are limited. Additionally, most reviewed models were evaluated on controlled research datasets, and performance often declines when applied to real-world data from different imaging equipment or institutions, highlighting the importance of evaluating generalizability before clinical deployment.
Future research should prioritize developing deep learning models that are more data-efficient and resource-conscious, enabling their use in low-resource clinical environments where large datasets and powerful computing infrastructure are unavailable. Semi-supervised and self-supervised approaches that can learn from unlabeled data represent promising directions.
A major gap in the literature is the lack of multiclass classification models that can distinguish between multiple breast cancer subtypes rather than simply separating benign from malignant cases. Clinical decision-making often requires finer-grained classification, and building reliable multi-category AI models is a key research priority.
Standardization of experimental evaluation is urgently needed. The variability in preprocessing techniques, dataset sizes, and evaluation metrics across studies makes direct comparison difficult and undermines reproducibility. Establishing consistent benchmarks and reporting standards would accelerate progress and help translate research findings into reliable, clinically deployable diagnostic tools.