Breast cancer became the most commonly diagnosed cancer worldwide in 2020, surpassing lung cancer for the first time. Imaging detection - including mammography, ultrasound, MRI, and PET - is central to screening, staging, and monitoring treatment response. However, the sheer volume of images generated by modern screening programs creates a crushing workload for radiologists, contributing to errors and delays in reporting.
Images with low quality or ambiguous features push radiologists to their diagnostic limits, and the same image interpreted by different specialists at different institutions can yield different conclusions. Computer-aided diagnosis (CAD) systems aim to standardize and automate interpretation, reducing radiologist workload and improving accuracy - but early CAD systems had high false-positive rates that limited their clinical value.
The arrival of deep learning (DL) - particularly convolutional neural networks - has transformed CAD capabilities. Unlike rule-based systems, deep learning algorithms learn diagnostic features directly from large collections of labeled images, enabling end-to-end learning that integrates image acquisition, feature extraction, and classification in a single optimizable pipeline.
This review covers AI applications across all four major breast imaging modalities - mammography, ultrasound, MRI, and PET - as well as the emerging fields of radiomics and radiogenomics, which use AI to extract molecular and genomic insights from imaging data without requiring invasive biopsies.
Machine learning (ML) is a subfield of AI that builds analytical models from labeled training data. ML algorithms broadly divide into supervised methods (classification and regression tasks with labeled examples), unsupervised methods (clustering and dimensionality reduction for unlabeled data), and reinforcement learning (self-improving systems that optimize decisions based on feedback). Traditional ML requires human experts to define which features of an image are diagnostically meaningful.
Deep learning (DL) eliminates this requirement by learning its own feature representations directly from raw image data. Inspired by the biological structure of the visual cortex, DL uses many layers of interconnected processing units. Early layers detect simple patterns like edges and textures; deeper layers recognize complex clinical structures like tumor margins and calcification patterns. This hierarchical approach closely mirrors how human radiologists process visual information.
The most widely used DL architecture for breast imaging is the convolutional neural network (CNN). CNNs use specialized convolution operations that efficiently detect spatial patterns across an image - critically, they are robust to small variations in the position or orientation of a finding, an important property when searching for tumors that can appear anywhere in the breast.
Two practical strategies that improve DL performance on medical images are transfer learning (adapting a network pre-trained on millions of natural images to medical imaging tasks) and data augmentation (artificially expanding limited training datasets through rotations, flips, and other transformations). Both strategies help overcome the fundamental challenge that labeled medical images are scarce and expensive to produce.
Mammography is the only imaging tool proven to reduce breast cancer mortality, with screened women showing a 53% lower risk of breast cancer-related death. CNN-based AI has been applied to mammography for four main tasks: benign vs. malignant tumor classification, mass localization, tissue segmentation, and breast density classification for risk stratification.
VGG16-based models have achieved high accuracy on mammography classification, outperforming earlier architectures like AlexNet, GoogleNet, and EfficientNet on the same task. However, CNN-based mass localization struggles in dense breast tissue and near the pectoral muscle - regions where tumor signals are obscured by overlapping high-density tissue, creating a fundamental challenge that deep learning cannot fully overcome without additional imaging information.
Contrast-enhanced spectral mammography (CESM) is an emerging technique that uses an iodine contrast agent to reveal tumor neovascularization. AI systems applied to CESM images have demonstrated the ability to reduce false-positive rates and, in one study, significantly reduce the biopsy rate for benign tumors - an important outcome because unnecessary biopsies carry physical and psychological costs for patients.
Digital breast tomosynthesis (DBT) generates 3D reconstructed images from multiple angles, significantly improving sensitivity over standard 2D mammography, especially for small lesions hidden by overlapping tissue. DL-based algorithms applied to DBT data have produced maximum suspicion projection images that outperformed manual reading, though DBT-based AI remains technically more complex than standard mammography AI.
Breast ultrasound is widely used for cancer detection and biopsy guidance because it is radiation-free, portable, and cost-effective. For non-calcified small lesions that mammography misses, ultrasound has distinct advantages. CNN-based AI applied to ultrasound images has achieved 90% accuracy with 86% sensitivity and 96% specificity for benign vs. malignant classification.
A key challenge for ultrasound AI is operator dependence: the quality of ultrasound images depends heavily on the technique and experience of the sonographer performing the scan. This creates systematic variability in training data that does not exist for fixed-protocol imaging like mammography, causing AI recognition of ultrasound images to lag behind CT and MRI AI development.
Advanced approaches have addressed this challenge through contextual adversarial learning frameworks for tumor segmentation, and through vision-transformer-style architectures like ConvNeXt that provide visual interpretations alongside malignancy score predictions. One study achieved 99.1% breast cancer classification accuracy by combining DarkNet53 feature extraction with evolutionary optimization algorithms for feature selection.
Ultrasound elastography - which measures tissue stiffness - has further improved AI-assisted diagnosis. A two-layer DL model applied to shear wave elastography images achieved 93.4% accuracy, 88.6% sensitivity, and 97.1% specificity for classifying tumor properties. AI analysis of combined B-mode ultrasound and elastography features can also predict axillary lymph node metastasis before surgery - important information for treatment planning.
Breast MRI is the most sensitive imaging tool for staging and high-risk screening, and its sensitivity is not affected by breast tissue density - a key advantage over mammography. MRI provides unique functional information including spectroscopic metabolite analysis, diffusion-weighted imaging of tissue microstructure, and pharmacokinetic mapping of blood flow patterns within tumors.
AI has been applied to MRI for pathological complete response (pCR) prediction - determining before surgery whether a tumor has been completely eliminated by neoadjuvant chemotherapy. This is clinically significant because pCR is one of the strongest predictors of long-term survival, and AI-based prediction could allow earlier adjustment of treatment regimens for non-responders.
Multiparametric MRI approaches combine diffusion-weighted and contrast-enhanced sequences to extract four types of features: tumor morphology, texture, hemodynamics, and pharmacokinetics. AI integration of these complementary features provides a more complete picture than any single MRI sequence. One research group integrated DL into picture archiving and communication systems (PACS), allowing radiologists to quickly identify tumor-containing slices instead of manually scrolling through hundreds of images.
MRI-based AI performance varies meaningfully across breast cancer subtypes. Studies indicate that MRI is better suited for evaluating neoadjuvant therapy response in HER2-positive and triple-negative breast cancer compared to luminal subtypes, reflecting biologically different patterns of tumor response that AI can be trained to recognize. Multimodal MRI combining multiple imaging sequences continues to outperform any single approach.
PET/CT (positron emission tomography with computed tomography) provides metabolic information by measuring how cancer cells consume glucose - a unique capability unavailable from structural imaging alone. 18F-FDG uptake correlates with tumor grade and proliferation rate, allowing AI models to extract prognostic information from PET scans. Dedicated breast PET has shown higher sensitivity and specificity than standard PET/CT for early breast cancer detection.
PET/MRI combines the sensitivity of MRI with the metabolic information of PET, achieving 90-99% sensitivity for breast cancer detection while providing lower radiation dose than PET/CT. AI applied to PET/MRI fusion images has further improved sensitivity and specificity for detection and staging compared to either modality alone. However, the high cost and limited availability of PET/MRI restrict its use to specific high-risk or complex clinical scenarios.
Radiomics uses AI to extract hundreds of quantitative features from imaging data - texture, entropy, morphology, and intensity distribution - that encode information about tumor biology at a level invisible to human observers. A radiomics model applied to MRI images from TCGA data successfully predicted molecular subtype classification including estrogen receptor, progesterone receptor, and HER2 expression status.
Radiogenomics takes this further by jointly analyzing radiomic imaging features and genomic data, establishing direct connections between how a tumor looks on a scan and which genes are expressed within it. This approach enables non-invasive molecular characterization of the entire tumor - overcoming the sampling limitations of needle biopsy, which captures only a small fragment of a potentially heterogeneous mass.
Despite strong benchmark results, AI-assisted breast cancer imaging faces several key barriers to routine clinical use. First, there is no standardized pipeline from image acquisition to diagnosis - different studies use different preprocessing steps, segmentation approaches, and feature extraction methods, making it impossible to compare results across institutions or to deploy a model from one site at another without revalidation.
Novel algorithms with higher specificity are needed to handle the full diversity of image quality and patient characteristics encountered in real clinical practice. Current AI systems were often developed on carefully curated datasets that do not reflect the full range of image quality, patient age, breast density, and cancer subtype that radiologists encounter daily.
A systematic review found that in retrospective studies, AI systems were on average inferior to radiologists in interpretation accuracy - challenging the optimistic benchmark results from smaller controlled studies. No prospective randomized controlled trials have yet directly compared AI as a standalone interpretation system against radiologist performance. Such trials are essential for understanding the real-world clinical value of AI before deployment at scale.
The radiogenomics frontier, while promising, faces significant practical barriers: it requires large samples, expensive genomic sequencing, and powerful computational capabilities. Standardized protocols for imaging acquisition and feature extraction must be established before radiogenomics can transition from research to routine clinical practice.
AI-assisted imaging diagnosis represents a genuinely promising direction for breast cancer, offering the prospect of faster, more accurate, and more consistent diagnosis across imaging modalities. The most valuable near-term applications are automated screening to reduce radiologist workload, improved detection of subtle findings in difficult cases, and non-invasive treatment response monitoring that reduces the need for repeat biopsies.
Realizing this potential requires addressing three interconnected gaps: developing standardized operational procedures that ensure reproducible results from imaging through to diagnosis; building algorithms robust enough to handle the diverse image quality and patient characteristics of real clinical populations; and validating AI tools in prospective randomized trials with large, high-quality patient cohorts.
A breakthrough technical requirement is establishing a direct connection pathway between AI training databases and hospital information systems (HIS) or health insurance databases. This would enable continuous learning from real-world clinical outcomes, creating a self-improving diagnostic platform that becomes more accurate over time as it processes more cases.
The broader vision is an open, self-optimizing AI imaging platform that integrates imaging findings, molecular data, and clinical outcomes - moving beyond isolated diagnostic tasks toward comprehensive tumor characterization that supports individualized treatment decisions across the full cycle of breast cancer care from screening to follow-up.