AI-Driven Microcalcification Detection in Digital Mammography for Early Breast Cancer Diagnosis: A Scoping Review, Challenges, Limitations, and Future Perspectives

Abstract

Background: Microcalcifications (MCs) are among the earliest mammographic signs of breast cancer, yet their detection remains challenging due to small size, low contrast, and dense breast tissue. This scoping review synthesizes AI-driven methods for MC detection in digital mammography, focusing on three dimensions: comparative performance of deep learning (DL) versus traditional methods, the clinical impact of explainable artificial intelligence (XAI), and the role of synthetic data in addressing dataset limitations. Methods: Following PRISMA-ScR guidelines, we systematically searched seven databases for studies published between January 2000 and January 2026. Of 366 initial records, 72 peer-reviewed studies were included in the final synthesis. Results: DL architectures, particularly convolutional neural networks (CNNs), have generally reported higher diagnostic performance (accuracy up to 99.71% and Area Under the Curve (AUC) up to 0.998) than traditional machine learning methods, although direct comparisons are hindered by heterogeneous datasets and evaluation protocols. XAI techniques have yet to undergo rigorous validation in real-world clinical settings, with very low certainty of evidence regarding their impact on radiologists’ trust or workflow integration. Synthetic data generation mitigates some data scarcity and privacy constraints but introduces artifacts (e.g., checkerboard patterns in 39–46% of cases) that limit clinical realism. Conclusions: DL offers substantial promise for MC detection, but translation to clinical practice requires robust XAI validation, higher-quality synthetic data, and prospective studies on diverse, longitudinal datasets.

1. Introduction

Breast cancer develops when cells in the breast grow uncontrollably, usually forming a tumor. The abnormal growth can be felt as a palpable lump or visualized on mammograms. If left untreated, cancer cells can spread to other parts of the body, resulting in metastasis. The presence of metastasis makes the cancer more difficult to treat [1].

Breast cancer is a public health challenge with high incidence and mortality. Breast cancer in women constitutes 32% of all cases according to the International Agency for Research on Cancer (IARC) [2,3].

Early detection enables timely intervention. It improves prognosis and reduces mortality rates. In this regard, Microcalcifications (MCs) are considered one of the earliestindicators of breast abnormalities [4].

MCs are small calcium deposits in breast tissue ranging from 0.1 to 1 mm. MCs can be benign or malignant. Each MC has a distinct shape and distribution pattern that has clinical significance. Benign MCs are typically large, round, smooth, and scattered. Malignant MCs are typically small, fine, clustered, and often require magnification for visualization [5,6].

Microcalcification Clusters (MCCs) are groups of three or more MCs within a one-square-centimeter area and are among the earliest indicators of breast cancer [5,7]. Regular screening tests, such as mammography, are designed to detect breast cancer before symptoms manifest [8].

MCCs are often associated with Ductal Carcinoma in Situ (DCIS) or early-stage invasive cancers. Accurate detection is critical for timely diagnosis and treatment. However, MCCs are difficult to detect because the constituent MCs are very small and exhibit low conspicuity. These challenges are especially severe in dense breasts [7].

Despite its effectiveness, traditional mammography may fail to detect low-contrast MCCs, resulting in false negatives. In addition, overlapping tissue structures and imaging artifacts can lead to false-positive findings. Therefore, unnecessary biopsies or missed diagnoses can result. Tools powered by artificial intelligence (AI) enhance MC detection. Machine Learning (ML) and Deep Learning (DL) algorithms are effective at identifying subtle patterns that humans may miss, thereby reducing diagnostic errors [9,10,11].

AI enhances MCC detection and supports earlier breast cancer diagnosis. AI adoption in breast cancer screening is a significant step forward in personalized medicine. It also reduces healthcare costs linked to late-stage treatments [12].

Despite the promising role of AI in improving MC detection, significant challenges remain, including model robustness, explainability to improve transparency, and data scarcity. In addition, the literature remains fragmented, with variations in datasets, evaluation protocols, XAI approaches, and clinical validation. Consequently, there is a need for a comprehensive synthesis of current AI-driven methods, their limitations, and the challenges that must be addressed to facilitate reliable adoption. This review provides a valuable resource for researchers, biomedical engineers, and healthcare technology developers by summarizing current advances, identifying current challenges, limitations, and future research directions.

Table 1. SPICE method used to formulate Research Question 1.

SPICE Element	Input	Scenario
S—Setting	Clinical context and primary modality	Early breast cancer detection via mammography.
P—Perspective	Research perspective	Researchers and developers.
I—Intervention	Deep learning approaches	Development and application of DL models, particularly Convolutional Neural Networks (CNN).
C—Comparison	Conventional approaches	Traditional ML methods and rule-based systems.
E—Evaluation	Performance metrics	Sensitivity, specificity, generalizability, and the Area Under the Curve (AUC).

,

Table 2. SPICE method used to formulate Research Question 2.

SPICE Element	Input	Scenario
S—Setting	Transition from research to clinical practice	Clinical implementation of AI for the detection of MCs and MCCs.
P—Perspective	Healthcare stakeholders	Radiologists, hospital administrators, and workflow specialists concerned with clinical integration and usability.
I—Intervention	Explainability in DL systems	Integration of explainable AI (XAI) techniques into deep learning models.
C—Comparison	Standard DL approaches	Black-box deep learning models versus XAI-enhanced models.
E—Evaluation	Human and operational outcomes	Radiologist trust, diagnostic confidence, workflow efficiency, and clinical adoption.

and

Table 3. SPICE method used to formulate Research Question 3.

SPICE Element	Input	Scenario
S—Setting	Data-centric challenges in AI	Medical imaging contexts with limited, imbalanced, or privacy-constrained datasets.
P—Perspective	Technical perspective	Data scientists and ML researchers focusing on model robustness and generalization.
I—Intervention	Synthetic data generation	Use of synthetic data to augment training datasets.
C—Comparison	Training strategies	Models trained on real data versus hybrid datasets (real + synthetic).
E—Evaluation	Model performance outcomes	Performance, generalizability across different clinical settings, and reduction in dataset bias.

present the Setting, Perspective, Intervention, Comparison, and Evaluation (SPICE) framework used to formulate the three research questions that guide this review and help identify current challenges, limitations, and future research directions.

Research Question 1: What is the performance of DL architectures, specifically Convolutional Neural Networks (CNNs), compared with that of traditional ML methods for the detection of MCs and MCCs in mammograms?

Research Question 2: To what extent does incorporating Explainable Artificial Intelligence (XAI) into DL-based detection systems for MCs improve radiologists’ diagnostic confidence in a clinical screening environment?

Research Question 3: To what extent does synthetic data generation mitigate dataset limitations (size, diversity, privacy) in training DL models for MC or MCC detection, and how does it impact model robustness and generalizability across heterogeneous clinical settings?

The objective of this paper is to present a scoping review of three key AI-enabled advances in the detection of MCs and MCCs in digital mammography for early breast cancer diagnosis: (1) the comparative diagnostic performance of DL approaches compared to conventional ML and rule-based methods; (2) the role and impact of XAI methods in improving clinical interpretability and workflow integration; and (3) the role of synthetic data generation in overcoming dataset limitations and improving model robustness and generalizability across heterogeneous clinical settings.

Distinction from Prior Reviews

Several reviews have examined AI-based methods for breast cancer diagnosis in mammography. Qureshi et al. [13] presented a comprehensive chronological review of breast cancer detection methods, spanning from traditional image processing techniques to DL (1970–2023). However, their review focused on general breast cancer detection rather than the specific analysis of MCs. Furthermore, the review did not address emerging topics such as XAI or synthetic data generation. Bashir and Bhosle [14] focused specifically on MC segmentation and classification. Nevertheless, their review did not extend to explainability or data augmentation strategies beyond traditional techniques. Añez et al. [15] combined a review with experimental validation and implemented XAI techniques such as the Gradient-weighted Class Activation Mapping (Grad-CAM) and the SHapley Additive exPlanations (SHAP). However, they treated these techniques as a pipeline component rather than critically examining their clinical evidence gap. Moreover, their review focused on general breast cancer detection rather than the specific analysis of MCs. From a clinical perspective, van Leeuwen et al. [16] conducted a meta-analysis on MC morphology in Ductal Carcinoma In Situ (DCIS), but their work did not incorporate computational methods.

To our knowledge, no prior review has simultaneously addressed three critical dimensions: (1) comparative DL performance against traditional methods, (2) the evidence gap in XAI clinical impact, and (3) synthetic data as a solution to dataset limitations. Furthermore, existing reviews rarely employ a formal SPICE framework to derive Research Questions or provide domain-specific risk of bias assessments for XAI and synthetic data.

Table 4. Comparative analysis of the present review with previous reviews on AI-based methods for MC detection.

Dimension	Qureshi et al. [13]	Bashir and Bhosle [14]	Añez et al. [15]	van Leeuwen et al. [16]	Present Review
Scope	ML, DL	MC seg.	Pipeline	Clinical	DL vs. ML + XAI + Synthetic
AI focus	Broad	MC	Pipeline	None	MC detection
PRISMA	Yes	No	Yes	Yes	Yes
SPICE/RQs	No	No	No	No	Yes
XAI	None	None	Implemented	None	Evidence gap analysis
Synthetic data	None	None	None	None	Yes
Risk of bias	–	–	PROBAST	QUIPS	Domain-specific
N studies integration	150+	–	–	Meta	72
Longitudinal studies	No	No	No	No	Gap identified

Legend: Bold = distinctive contribution of the present review.

highlights the distinguishing features of the present study.

The contributions of this review are as follows.

• Comparative analysis of detection methods: A review of DL, ML, and rule-based approaches for the detection of MCs and MCCs, highlighting their performance, limitations, and clinical applicability.
• Explainability in clinical AI systems: An analysis of the role of XAI techniques in enhancing interpretability, radiologists’ trust, and workflow integration in AI-assisted mammography systems.
• Data-centric challenges and solutions: A review of synthetic data generation techniques to mitigate challenges posed by limited, imbalanced, or privacy-constrained datasets, and their impact on model robustness, reduction in dataset bias, and generalizability across heterogeneous clinical settings.

Longitudinal screening was not addressed in this review because the available literature was predominantly cross-sectional and focused on single-time-point detection. Nevertheless, the lack of longitudinal studies represents an important research gap that warrants further investigation.

The remainder of this paper is structured as follows. Section 2 describes the methodology, including the search strategy, study selection process, data extraction, and synthesis approach following PRISMA-ScR guidelines. Section 3 provides a clinical perspective on why microcalcification detection remains challenging. Section 4 reviews the evolution of detection methods, from visual inspection and rule-based systems to traditional machine learning and deep learning architectures, with a critical comparison of their strengths and limitations. Section 5 examines the role of the datasets in advancing AI-driven MC detection. Section 6 synthesizes the key challenges and limitations identified in the literature, including issues of data quality, model robustness, explainability, and synthetic data generation. Section 7 presents the synthesis of the three research questions guiding this review. Section 8 assesses the risk of bias due to missing results, and Section 9 discusses the certainty of evidence. Finally, Section 10 outlines future research directions and concludes the review.

2. Methodology

This paper follows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) [17] guidelines while maintaining key elements of PRISMA 2020 [18]. A total of 72 peer-reviewed studies published between January 2000 and January 2026 were synthesized. These studies cover approaches ranging from rule-based systems to current DL architectures for MC and MCC detection.

Visual inspection methods were included as baseline approaches to contextualize the transition to rule-based and AI-driven systems. This scoping review was not registered in a review registry, and no review protocol was developed prior to study initiation.

2.1. Search Strategy

A literature search was conducted in electronic databases, including Web of Science, Scopus, PubMed, IEEE Xplore, ScienceDirect, SpringerLink, and Google Scholar. Search strings were adapted to the syntax and indexing requirements of each database to ensure consistency and reproducibility.

For Google Scholar, the first 50 results ranked by relevance at the time of the search were screened to identify additional relevant studies. The search strategy targeted studies addressing the detection, classification, diagnosis, or segmentation of MCs or MCCs using visual inspection, rule-based systems, or ML/DL approaches.

The search components are summarized in

Table 5. Search strategy, inclusion and exclusion criteria. * represents a wildcard character.

Component	Details
Objective	To identify studies on visual inspection, rule-based systems, and ML/DL methods for MC detection in digital mammography for early breast cancer detection.
Databases	Web of Science, Scopus, PubMed, IEEE Xplore, ScienceDirect, SpringerLink, and Google Scholar
Timeframe	January 2000 to January 2026
Search Concepts	Population/Setting: Digital mammography, breast imaging, breast screening Target: Microcalcifications, calcification clusters Method: Visual inspection, rule-based systems, machine learning, deep learning, artificial intelligence Outcome: Detection, classification, diagnosis, segmentation
Search Terms	Mammography: mammogra*, mammogram*, “breast imaging”, “breast screen*”, “digital mammogra*” Microcalcifications: microcalcif*, calcif*, MC, MCC, “microcalcification cluster*”, “breast calcif*” Methods: “visual inspection”, “rule-based”, “machine learning”, “deep learning”, “artificial intelligence”, “neural network*”, “CNN”, “SVM”, “random forest”, “k-NN”, “XAI”, “interpretability”, “synthetic data” Detection: detect*, classif*, diagnos*, segment*, identif*, “computer-aided detection”, CAD
Boolean Logic	(mammogra* OR mammogram* OR “breast imaging”) AND (microcalcif* OR calcif* OR MC OR MCC OR “calcification cluster*”) AND (“visual inspection” OR “rule-based” OR “machine learning” OR “deep learning” OR “artificial intelligence” OR AI OR “computer-aided detection” OR CNN OR SVM OR “random forest”) AND (detect* OR classif* OR diagnos* OR segment*) AND (“explainable AI” OR “XAI” OR “interpretability” OR “synthetic data” OR “data augmentation”)
Inclusion Criteria	Studies focusing on detection, classification, or segmentation of microcalcifications (MCs) or microcalcification clusters (MCCs) in digital mammography Studies employing visual inspection, rule-based methods, ML/DL, or XAI techniques Studies evaluating synthetic data or data augmentation strategies in relevant models Original peer-reviewed research articles with sufficient methodological detail Studies reporting quantitative or qualitative evaluation metrics English-language publications with full text available
Exclusion Criteria	Studies focusing exclusively on clinical outcomes without methodological details General ML/DL studies without application to MCs Studies using only non-mammography imaging modalities (e.g., ultrasound, MRI) Conference abstracts, editorials, letters, and non-English publications
Additional Search	Manual screening of reference lists (backward snowballing)

. Key elements included the population/setting (digital mammography), target lesions (MCs and MCCs), AI methodologies (rule-based, ML/DL, and XAI-related approaches), and outcomes (detection, classification, diagnosis, and segmentation).

Inclusion criteria were peer-reviewed original research articles that focused on the detection of MCs and MCCs in digital mammography. Studies had to present quantitative or qualitative evaluation metrics, enough methodological detail to be reproducible and be available in English in full-text.

Exclusion criteria included studies not focused on MC detection, studies using imaging modalities other than mammography, conference abstracts, editorials, and letters lacking full methodological descriptions, as well as non-English publications.

In addition, backward snowballing was performed by manually screening the reference lists of included studies to identify further relevant articles.

2.2. Selection Process

Table 6. Screening process and results.

Component	Details
Identification	366 records identified; 31 duplicates records removed prior to screening
Screening Phase 1: Title/Abstract	Records screened: 335 Records excluded: 171 Exclusion breakdown: - Not focused on MC detection or involving irrelevant applications: 73 - Not ML/DL-based: 61 - Wrong imaging modality: 37 Records remaining for full-text assessment: 164
Screening Phase 2: Full-Text Assessment	Full-text articles assessed: 164 Full-text articles excluded: 92 - Focused solely on clinical outcomes (survival rates, treatment efficacy): 15 - General ML/DL studies without specific MC application: 10 - Studies with vague or absent contribution statements: 7 - Studies with no quantitative results: 15 - Full-text articles not accessible: 8 - No original research on MC detection: 15 - Conference abstracts without full methodology: 12 - Non-English publications: 10 Articles meeting inclusion criteria: 72
Final Included Studies	Total included: 72 articles. Primary research studies: 66 (91.7%)—Original research on MC detection methods, including visual methods, rule-based methods, quantitative results, key findings, and ML-based methods. Review/foundational studies: 6 (8.3%)—Systematic reviews, meta-analyses, and theoretical papers.
Contextual References	Additional documents: 9 Sources: Selected for background narrative only and not included in the synthesis or data extraction. Purpose: To provide background context for the introduction, factors affecting MC detection, visual methods, and datasets used in MC detection research.

summarizes the screening process and its results. A total of 366 records were identified through database searches. After the software-assisted removal of 31 duplicate records, titles and abstracts were independently screened by two reviewers. Disagreements were resolved by a third reviewer.

The screening followed predefined inclusion and exclusion criteria described in Table 5. After title and abstract screening, 171 records were excluded due to a lack of relevance to MC detection in digital mammography, focus on non-relevant breast lesions, use of non-mammography imaging modalities, and where applicable, absence of relevant computational methods aligned with the review scope. Full-text articles were then assessed for eligibility. Studies that did not meet the inclusion criteria were excluded, and the reasons for exclusion were documented.

Of the 164 full-text articles assessed for eligibility, 92 were excluded. The main reasons for exclusion were a primary focus on clinical outcomes, such as survival rates or treatment efficacy (n = 15); general ML or DL studies without a specific application to MC detection (n = 10); unclear or absent statements of contribution (n = 7); lack of quantitative performance results (n = 15); inaccessible full-text articles (n = 8); absence of original research on MC detection (n = 15); conference abstracts lacking sufficient methodological detail (n = 12); and publications not written in English (n = 10).

The screening process resulted in 72 studies being included in the final synthesis. Additionally, nine contextual references were retained to provide background information and support the interpretation of the findings. These references were not included in the formal evidence synthesis and were used exclusively to establish narrative context and discuss relevant developments in the field.

Table 7. Contextual References Supporting the Narrative (n = 9).

Component	Description
Contextual documents	9
Sources	Selected primary and foundational studies from the 72 included articles, along with relevant background papers from the 92 articles excluded during the full-text review.
Purpose	Provide clinical and methodological context; support the Section 1, the Factors affecting the MC detection sections, and frame the research problem. Provide definitions, ML/DL mathematics, models, advancements in the classic ML.

lists the contextual references and their purpose.

2.3. Risk of Bias

The methodological characteristics of the included studies were extracted to support the narrative synthesis and interpretation of findings. A formal meta-analysis was not conducted due to substantial heterogeneity in study designs, datasets, and performance metrics. As part of a descriptive assessment of the included literature, the following characteristics were recorded for each primary study:

• Study Design: Retrospective analysis, prospective validation, or comparative bench-marking.
• Data Source: Public versus private datasets; screening versus diagnostic data.
• Validation Strategy: Internal validation (e.g., cross-validation) or external validation on independent datasets.
• Reference Standard: Method used to establish the ground truth (e.g., biopsy confirmation, expert radiologist consensus).

This descriptive assessment summarizes key methodological characteristics and po-tential sources of bias. It supports the interpretation of results across studies. Due to study heterogeneity and variability in reported outcomes, a formal risk-of-bias assessment tool for diagnostic accuracy studies (e.g., QUADAS-2) was not applied.

2.4. Data Extraction

A standardized data extraction form was used to capture key information from all included studies. Two reviewers independently extracted the data. Discrepancies were resolved through discussion or consultation with a third reviewer. Extracted information included:

• Bibliographic Information: Authors, publication year, and study title.
• Study Characteristics: Primary task (detection, segmentation, classification, or diagnosis), methodological approach, and specific algorithm(s) used.
• Dataset Information: Name and type of dataset(s) used (e.g., INbreast, DDSM), number of images or cases, and key dataset characteristics mentioned.
• Key Findings: Main outcomes, reported performance metrics (e.g., sensitivity, specificity, AUC), and principal conclusions.
• Limitations and Future Work: Challenges, limitations, and suggested future directions reported by the study authors.

Quantitative performance metrics were recorded as reported by the original authors. For studies evaluating multiple models, data for the final model or the model emphasized as primary by the authors were extracted. Figure 1 shows the PRISMA 2020 [18] flow diagram for the study identification, screening, and inclusion process generated with [19].

2.5. Data Synthesis

Due to substantial heterogeneity in study designs, methodologies, datasets, and performance metrics, a quantitative meta-analysis was not appropriate. Instead, a structured qualitative synthesis was conducted [17].

Data were analyzed thematically according to a predefined analytical framework aligned with the objectives of the review:

• Evolution of Technical Approaches: Charting the progression from visual inspection and rule-based systems to traditional ML and contemporary DL architectures.
• Data Ecosystem and Its Influence: Analyzing the role, characteristics, and reported limitations of datasets used in model development and validation.
• Reported Challenges and Limitations: Synthesizing the technical, clinical, and translational barriers identified across the literature.
• Emerging Trends and Proposed Solutions: Collating future directions and novel approaches proposed by researchers.

Findings within each theme were synthesized narratively. Supporting tables were used to summarize study characteristics and key methodological concepts.

2.6. Data Presentation and Summary

The results of the synthesis are presented in narrative form, following the analytical framework described above to enhance clarity and accessibility, as follows:

• Summary tables (e.g.,

  Table 8. Comparison of MC detection approaches in mammography, highlighting methodological paradigms, feature dependency, performance behavior, and key limitations.

  Method	Paradigm	Feature Dependency	Strengths	Key Limitations
  Visual Inspection	Human expert analysis.	No explicit feature extraction.	Integrates clinical context; adaptable to complex cases.	Subjective; inter-observer variability; fatigue-related errors.
  Rule-based Systems	Deterministic/heuristic.	Handcrafted image features.	Interpretable; clinically trans-parent workflow.	Poor generalization; limited adapt-ability to imaging variability.
  SVM	Classical machine learning.	Strong dependence on engineered features.	Effective in high-dimensional separable spaces.	Sensitive to noise; limited modeling of complex morphology.
  Random Forest	Ensemble learning.	Moderate feature engineering required.	High accuracy; reduced over-fitting.	Limited interpretability; weaker performance on complex patterns.
  k-NN	Instance-based learning.	Distance-based feature space dependency.	Simple; effective with well-designed descriptors.	High computational cost; sensitive to dimensionality and irrelevant features.
  ANNs	Shallow neural learning.	Learned features (limited depth).	Improved adaptability over classical ML.	Overfitting risk; limited robust-ness across imaging domains.
  Decision Trees	Rule-based learning.	Moderate feature dependence.	Highly interpretable decision paths.	Overfitting; limited scalability to complex imaging data.
  Naive Bayes	Probabilistic learning.	Strong independence assumption.	Fast; computationally efficient.	Poor modeling of feature correlations in imaging data.
  XGBoost	Boosted ensemble learning.	Feature-engineered or hybrid features.	High predictive performance; strong non-linear modeling.	Reduced interpretability; sensitive to dataset bias and feature quality.
  Deep Learning	Representation learning (CNN-based).	Minimal manual feature engineering.	State-of-the-art accuracy; automatic hierarchical feature extraction; strong performance on MC/MCC patterns.	Requires large annotated datasets; domain shift sensitivity; limited interpretability; false positives in dense tissue.

  ) are used to present comparative overviews of key concepts.
• Descriptive statistics (e.g., counts and percentages of studies by methodology type) are provided where they meaningfully illustrate the scope of the evidence.
• The narrative integrates direct findings from the extracted data to describe the state of the field, identify consensus points, and highlight areas of contradiction or evidence gaps.

The included studies provide a structured basis for examining factors that influence the detection of MCs, comparing methodological approaches, and assessing the datasets used in AI-driven mammography research. This synthesis highlights key limitations in the current literature and identifies potential directions for future research in automated MC detection.

3. Why MC Detection Remains Challenging: A Clinical Perspective

The difficulty of detecting MCs in mammography arises from a combination of biological, technical, and human factors. These factors motivate the need for AI-based assistance.

3.1. Morphological and Distributional Complexity

MCs exhibit diverse shapes with clinical significance. Amorphous/indistinct calcifications may represent early malignancy or benign processes. Fine pleomorphic and fine linear forms are more suspicious for DCIS [5]. Distribution patterns also matter. Linear and segmental distributions raise higher suspicion than regional or scattered patterns [5].

3.2. Technical Imaging Factors

Low contrast, suboptimal spatial resolution, and detector-generated noise can reduce the visibility of MCs, particularly when the calcifications are small or faint [20]. Further-more, high-resolution imaging may amplify image noise, complicating the differentiation between true MCs and imaging artifacts [21].

3.3. Breast Density and Anatomical Overlap

Dense fibroglandular tissue can obscure MCs, reducing mammographic sensitivity and increasing the false-negative rate. Younger women are especially affected [22]. Overlapping anatomical structures can mimic or hide MCs. Examples include vessels and ducts [23].

3.4. Human Factors

Detection accuracy depends on radiologists’ experience. High workloads can lead to fatigue-related errors [24].

4. MC Detection Methods

Detection methods for MCs are essential for breast health, as they enable the identification of small calcium deposits in breast tissue that may indicate underlying abnormalities.

4.1. Visual Inspection

Manual examination of mammograms by trained radiologists has long been considered the gold standard for MC and other breast abnormality detection. Radiologists rely on their expertise to identify subtle patterns, shapes, and densities that may indicate the presence of benign or malignant lesions. However, this task is particularly challenging, as MCs are often small, faint, and sparsely distributed, requiring meticulous visual attention.

Observer variability can lead to inconsistent interpretations and diagnostic outcomes [25]. In addition, the evaluation of large volumes of mammograms introduces fatigue, which may reduce diagnostic accuracy and increase the likelihood of missed detections or false positives [26]. These limitations highlight the need for AI-assisted tools to support radiologists in achieving more consistent and accurate assessments [27,28,29,30,31,32].

Visual inspection depends entirely on human expertise, whereas computational approaches, including rule-based systems and ML/DL models, aim to improve the detection process by automatically identifying suspicious regions in mammograms.

4.2. Rule-Based Systems

Rule-based systems rely on handcrafted rules derived from morphological, intensity, and texture characteristics, together with classical algorithms such as spatial-domain methods, transform-based methods, and fuzzy logic, to identify suspicious MC regions.

The first applications of rule-based systems were developed using the Mammographic Image Analysis Society (MIAS) database [33,34]. These systems followed a pipeline of preprocessing, segmentation, feature extraction, and classification. Preprocessing used filters for noise reduction and enhancement. Segmentation relied mainly on thresholding rules and morphological operations to isolate potential MCs [35,36,37]. Feature extraction then applied rules based on size, shape, and intensity, and classification was carried out using heuristic rules to distinguish between MCs, noise, and normal tissue.

Some studies have reported high accuracies in detection, with sensitivity of up to 96.5% and accuracy above 90%. However, this accuracy can vary based on factors such as the type of transformation used, image quality, and the complexity of the rule-based system [33,38,39,40,41,42,43,44,45,46,47].

The main advantage of rule-based systems is their explainability. They are transparent in decision-making. This is especially valuable in clinical settings because interpretability is crucial in these settings [48].

However, rule-based systems depend on predefined heuristics and thresholds that may not generalize well across diverse datasets and imaging conditions, limiting their performance. In contrast, CNNs automatically learn relevant features and patterns from data without relying on manually defined rules [33,49].

Early rule-based systems were developed and validated almost exclusively using Screen Film Mammography (SFM) datasets. The MIAS [34] and DDSM [50] databases were particularly common. However, film mammography has inherent limitations compared to modern Full-Field Digital Mammography (FFDM). These limitations include lower contrast resolution, fixed dynamic range, and the need for digitization before computational analysis. These characteristics influenced the design of rule-based systems that relied on handcrafted features. The features were adapted to the noise and resolution properties of digitized films. Consequently, rule-based systems that performed well on SFM datasets often showed degraded performance when applied to FFDM images. This highlights the importance of understanding the underlying imaging modality.

4.3. Machine Learning-Based Methods

The detection of MCs in mammography has advanced due to the integration of ML techniques that have the potential to improve breast cancer diagnosis. Traditional ML-based methods have been widely used for early detection of breast cancer. Support Vector Machines (SVMs) were commonly employed [51,52,53,54,55]. SVMs can handle high-dimensional data to separate benign from malignant MCs using an optimal hyperplane. However, they are sensitive to noise and also require extensive feature selection [56,57]. Furthermore, challenges in morphological analysis hinder SVM robustness [58]. Nevertheless, integrating SVM with novel techniques remains critical for enhancing early breast cancer detection [57,59]. In contrast, DL techniques offer more advancements to detect MCs [60].

For a binary classification task (benign versus malignant MCs), a classifier f: R^d → {0, +1} is learned from a set of training examples (x_i, y_i)^N, where x_i ∈ R^d denotes the feature vector and y_i ∈ {0, +1} denotes the corresponding class label.

In SVM, y_i ∈ {−1, +1} and the decision function is given by f (x) = sign(w⊤ ϕ(x) + b), where w is the weight vector, b is the bias term, and ϕ(·) maps the input features into a higher-dimensional feature space. The classifier assigns a sample to the positive class if w⊤ ϕ(x) + b > 0 and to the negative class otherwise.

Random Forest (RF) is an ensemble learning method. It builds many Decision Trees (DT) during training. RF algorithms perform well in classifying MCs [51,52]. They achieve high accuracy due to ensemble learning [61] and computational efficiency. However, RF algorithms cannot capture fine-grained spatial relationships needed for MCC interpretation [62]. Their limited interpretability is also a critical limitation in clinical settings [63]. In RF, f is an ensemble of M decision trees T₁, …, T_M. The final prediction is obtained through majority voting: f (x) = majority vote{T₁(x), …, T_M(x)}.

The k-Nearest Neighbors (k-NN) algorithm is used for MC detection [52,64]. It relies on texture-based features [65]. This improves accuracy and efficiency. k-NN works well for local patterns. However, performance drops on large datasets. It also cannot model spatial clustering needed for MCC detection [56]. k-NN classifies a query point x by majority vote among its k nearest neighbors, that is, $\widehat{y}$(x) = mode{y_i: i ∈ N_k(x)}, with N_k(x) defined by a distance metric d(x, x_i) (e.g., Euclidean).

Artificial neural networks (ANNs) improved detection performance by learning non-linear feature relationships [56]. However, they are sensitive to noise and depend on large, diverse datasets [57,66]. Classical ANNs lack mechanisms for spatial aggregation of MCs into MCCs. This reduces effectiveness for cluster-level interpretation [59]. ANNs interpretability is also a concern. Clinicians may hesitate to trust “black-box” models for critical decisions [59]. Simple ANNs offer valuable support. However, their limitations call for more advanced techniques that balance precision and clinical utility [67]. For binary classification, the output layer uses a sigmoid function: σ(z) = ${\displaystyle \frac{1}{1+e^{-z}}}$. This maps the logit z to a probability p ∈ [0, 1]. The model predicts the positive class when p > 0.5.

More advanced ML methods have also been explored. These include DTs, Naive Bayes (NB), and XGBoost. XGBoost is an ensemble learning method based on boosting with DT. XGBoost achieves strong classification performance [64,68]. However, it remains dependent on handcrafted features and does not inherently model spatial grouping. This makes it less effective for MCC characterization. XGBoost also faces limitations related to data bias, feature selection, and model interpretability [57]. DTs offer strong interpretability. However, DTs are prone to overfitting [52]. NB classifiers are computationally efficient. Nevertheless, they perform poorly when modeling complex data distributions [59]. XG-Boost has demonstrated superior performance and computational efficiency, making it particularly suitable for clinical mammography applications [61,64,68].

Recent studies emphasize the effectiveness of DL algorithms. DL is an ML technique that uses multi-layer neural networks to automatically learn hierarchical features from raw data. These algorithms can detect MCs and analyze their spatial organization into MCCs. This highlights the clinical relevance of accurately characterizing MC distributions [69,70]. However, DL models can misinterpret benign structures as malignant MCs and cause false positives [71]. Moreover, unlike transparent rule-based systems, DL architectures suffer from limited explainability, a drawback that undermines clinician trust and complicates regulatory approval [72].

DL enables effective discrimination between benign and malignant MCs [73]. The use of diverse and representative datasets is essential. This ensures robust model performance across varying imaging conditions. Advanced DL architectures, such as depthwise-separable CNNs, have also shown improved performance in MC prediction compared to traditional techniques. This improvement comes from their ability to efficiently capture fine-grained features while maintaining computational efficiency [74].

A different but highly competitive family of DL models is You Only Look Once (YOLO). It detects and classifies MCs. It performs real-time object detection with high accuracy. Recent approaches have demonstrated the effectiveness of YOLO-based modelsin identifying subtle MCs in mammograms [75]. The integration of image enhancement techniques with YOLOv4 improves the visibility of fine details and enhances detection performance [76]. More advanced variants, such as YOLOv7-based frameworks, have also been proposed. These frameworks enable semi-automatic detection and improve the localization and delineation of suspicious regions [77,78]. However, the standard downsampling operations within YOLO’s architecture discard fine-grained, low-level details. These details are critical for identifying MCs [78].

Recent studies show the robustness of DL techniques. An optimized CNN model significantly improves diagnostic accuracy for breast MCs [79]. This finding suggests that further integration into clinical workflows could enhance radiologists’ diagnostic confidence. Similarly, an ensemble DL framework tailored for automated MC detection has been shown to markedly improve early breast cancer diagnosis [80].

DL replaces handcrafted features with learned representations h(l) = σ(W(l)h(l − 1) + b(l)). Parameters are optimized via backpropagation to minimize a loss function such as binary cross-entropy of Equation (1).

$$L=-{\displaystyle \frac{1}{m}}\sum _{i=1}^m\left[y_i\log { \widehat{y}}_i+(1-y_i)\mathrm{l}\mathrm{o}\mathrm{g}(1-{\widehat{y}}_i\right)]$$

(1)

General-purpose CNNs do not necessarily achieve superior performance for MCC detection in mammograms. This suggests that general-purpose architectures may not be optimal for this task [11].

Instead, task-specific CNNs tailored to the characteristics of MCs can achieve high classification performance while reducing computational complexity [11]. Domain differences in mammograms remain a major challenge in medical imaging. DL models trained on one domain often generalize poorly to others [81,82].

DL has also been applied to enhance image quality. Super-resolution frameworks improve visual assessment by revealing subtle MCs that may be missed by conventional methods [83]. Furthermore, DL-based approaches have been used to model microcalcification clusters (MCCs) for predictive purposes, underscoring the importance of spatial context in diagnosis. Traditional techniques often overlook this spatial context [84].

Image processing techniques have played a fundamental role in the evolution of MC detection and form the basis of many ML approaches [67]. Early methods improved visibility via enhancement, segmentation, and feature extraction. Recent approaches use texture analysis for better classification [85]. Table 8 compares MC detection approaches in mammography, highlighting methodological paradigms, feature dependencies, performance behavior, and key limitations.

A critical distinction often overlooked in prior reviews is whether a method was designed specifically for MC detection or merely adapted from generic object detection or segmentation architectures. In this review, we prioritize and give higher visibility to methods that incorporate MC-specific domain knowledge, as they are more likely to generalize robustly and achieve clinically relevant performance.

Among the 72 studies reviewed, we identified that only 17 studies (24%) proposed architectures or combinations of architectures with explicit MC-specific innovations [10,11,70,76,77,79,86,87,88,89,90,91,92,93,94,95,96]. This indicates that the majority of studies (76%) still rely on generic off-the-shelf architectures adapted for MC detection. Therefore, there is substantial opportunity for further innovation in MC-specific design, particularly for challenging scenarios such as subtle MCs in dense breast tissue. In this review, we prioritize a detailed discussion of these specialized methods.

5. The Role of the Datasets

Publicly available datasets play a crucial role in advancing MC detection techniques. They are an essential resource for algorithm development and validation. Large, well-annotated datasets facilitate the effective training of DL models and improve their generalization performance. For example, Pesapane et al. [69] showed improved detection and classification with large-scale datasets. Loizidou et al. [97] highlighted how dataset availability supports automated MC detection via temporal subtraction.

Diaz–Huerta et al. [58] emphasized the importance of quantitative analysis and demonstrated that publicly available datasets promote the development of robust algorithms. Luna et al. [81,98] underlined the relevance of curated multi-resolution datasets. These help train DL models that handle cross-domain variability. Improved diagnostic performance also depends on large and diverse datasets. This was reported by Wang et al. [98] and Mahmood et al. [99]. Such datasets enhance the robustness of ML-based approaches.

Recent studies show that publicly accessible datasets are fundamental for intelligent detection systems [92,100]. Frameworks like the Automatic Segmentation and Classification of Breast MIcroCAlcifications from Mammograms (DeepMiCa) [89] and Computer-Aided Detection (CAD) systems [7] highlight the importance of annotated datasets for clinical deployment. Teoh et al. [80] reinforce the need for optimized datasets in ensemble learning frameworks. These findings underscore the critical role of high-quality datasets in enabling reliable AI systems for early breast cancer diagnosis.

The MIAS dataset [34] contains 322 mammographic images from 161 patients. Each image size is 1024 × 1024 pixels. The resolution is 50 µm per pixel. It includes normal, benign, and malignant cases with expert-annotated lesion centers and radii. Abnormalities cover MCs, masses, and distortions. MIAS is a foundational dataset for early rule-based and classical ML approaches.

The INbreast dataset [101] is a high-quality FFDM collection. It contains 410 mammograms from 115 patients acquired with Siemens systems. Images are in DICOM format at 70 µm per pixel resolution. This dataset provides pixel-level annotations for MCs, masses, and asymmetries validated by radiologists with the Breast Imaging-Reporting and Data System (BI-RADS) [5] and pathology labels.

The Breast Cancer Digital Repository (BCDR) [102] supports breast cancer diagnosis research. It has two subsets: BCDR-Film and BCDR-Digital. They contain SFM and FFDM images, respectively. The dataset provides clinical metadata, BI-RADS descriptors, and lesion annotations, including MCs and masses. These features enable research on domain adaptation and cross-modality learning. Consequently, BCDR is a valuable resource for studying the transition from analog to digital mammography.

The Digital Database for Screening Mammography (DDSM) [50] is a large public dataset for MC and mass detection. It contains about 2620 screening cases annotated by expert radiologists (lesion boundaries, pathology, malignancy). Each case includes craniocaudal (CC) and mediolateral oblique (MLO) views of both breasts. Images were acquired with SFM mammography and digitized at up to 43.5 µm per pixel resolution.

The Curated Breast Imaging Subset (CBIS-DDSM) [103] is a curated subset of the Digital Database for Screening Mammography (DDSM) [50]. It contains reprocessed and normalized mammographic images, corrected metadata, and predefined training and testing splits. Refined annotations support consistent experimental evaluation. It explicitly separates calcification and mass lesions, which is valuable for MC-focused studies. However, it is based on SFM-digitized mammography. Therefore, it lacks modern FFDM imaging characteristics.

The MEXBreast [97] dataset is a publicly available mammography database comprising 620 images acquired at multiple spatial resolutions (50, 70, and 100 µm). It is designed to support research in multi-domain MC detection and diagnosis, particularly within full-field digital mammography (FFDM) and DL-based CAD systems. The dataset includes both normal and abnormal cases, with expert annotations of MCs and associated BI-RADS categories.

Table 9. Comparison of publicly available mammography datasets for MC detection and classification, including limitations.

Dataset	Size/Patients	Imaging Type	Limitations
MIAS [34]	322 images/161 patients	SFM (digitized)	Very small scale; outdated screen-film modality; limited annotation precision (approximate lesion centers only); not representative of modern FFDM data.
INbreast [101]	410 images/115 patients	FFDM (70 μm)	Limited dataset size reduces generalization ability; class imbalance; single-center acquisition limits domain diversity.
BCDR [102]	Multiple sub-sets	SFM and FFDM	Moderate annotation granularity; variability between film and digital subsets; limited standardization compared to newer datasets.
DDSM [50]	∼2620 cases	SFM (digitized)	Old imaging modality (screen-film); digitization noise; inconsistent annotation quality; requires extensive preprocessing for DL.
CBIS-DDSM [103]	Curated subset of DDSM	SFM (digitized)	Still inherits limitations of DDSM; lacks FFDM characteristics; reduced variability due to preprocessing pipeline.
MEXBreast [97]	620 images	Multi-resolution FFDM (50, 70, 100 µm)	Limited public benchmarking adoption; relatively small scale compared to DDSM.

presents a comparison of publicly available mammography datasets for MC detection and classification.

6. Challenges and Limitations

Table 10. Comparative synthesis of included studies showing the maximum reported performance for rule-based, ML-based, and DL-based methods.

Metric	Rule-Based	ML-Based	DL-Based
Sensitivity	96.5% [46]	94.4% [55]	100% [29]
Accuracy	90.16% [55]	97.31% [62]	99.71% [11]
AUC	0.9677 [55]	0.9816 [62]	0.998 [94]

Note: Performances were reported across different studies and datasets. Variations in evaluation protocols and database characteristics may affect direct comparison between methods.

shows a comparative synthesis of the included studies, highlighting the maximum reported performance of rule-based, ML-based, and DL-based methods. Among the 72 reviewed papers, DL-based approaches reported higher performance in MC detection. Despite their promising performance, these methods still face several challenges and limitations. One of the primary concerns is data quality, as inaccuracies and inconsistencies in mammographic datasets can significantly affect model performance [104].

Noise in training datasets can lead to overfitting, resulting in poor generalization to unseen cases [88]. Although DL offers significant benefits, its adoption remains challenged by high computational demands and limited interpretability. Saliency maps and other post hoc explanations are unreliable. These are major obstacles for clinical use and regulatory approval. Variations in image protocols across mammographic scanners significantly impact model performance [81]. Bias in datasets skews model performance. It also raises ethical concerns about deploying models across diverse populations [105].

Breast density is another critical challenge. Dense tissue obscures MCs and increases false negatives [69]. Technological limitations affect DL efficacy. These include inadequate training data and limited interpretability. Optimized ensemble DL frameworks improve detection rates. However, they still struggle with image quality and patient variability [80]. High-resolution imaging with CNNs enhances accuracy. Still, robustness challenges persist [79].

Model robustness is also a critical issue. Models must perform consistently across diverse populations and imaging conditions. Nevertheless, breast density and acquisition variations complicate MC detection. Models need to generalize beyond controlled datasets [98]. Some architectures work well on specific datasets but fail with real-world variability [95].

Explainability is essential in clinical settings. Practitioners need clarity on model decisions to build trust [106,107]. DL model opacity limits clinical adoption. Various interpretability methods have been proposed to improve transparency. For example, radiomic signatures and explainability boost MC classification reliability [61]. Automatic segmentation improves interpretability by providing clearer visual representations [89]. However, XAI methods have not been integrated into routine clinical practice. Therefore, it is not possible to quantify their real clinical effect. Still, balancing interpretability and performance remains a critical challenge for viable systems [108,109].

A major limitation is the insufficient diversity of available data [93,110]. Generative models, such as generative adversarial networks (GANs), have emerged as a solution to overcome data limitations [111]. Generative adversarial networks (GANs) consist of a generator G and a discriminator D playing the minimax game of Equation (2).

$${min}_G{max}_D{E}_{\mathrm{x}~p_{data}}\left[\log D\left(x\right)\right]+E_{z~p_z}\left[log\left(1-D\left(G\left(z\right)\right)\right)\right]$$

(2)

where x denotes a real data sample, drawn from the training dataset, according to the distribution p_data, and z denotes a latent noise vector sampled from the distribution p_z. The generator G transforms z into a synthetic sample G(z), while the discriminator D estimates the probability that an input sample is real rather than generated.

GANs can synthesize realistic mammographic patches with MCs. However, their training is unstable [112]. Generated images often contain artifacts (see

Table 11. Limitations of mammogram simulation approaches.

Limitation	Description
GAN-Based Arti-facts	Checkerboard artifacts with a high incidence (39–46%). Breast boundary artifacts show the highest incidence (34–53%). Other issues include abnormal nipple-areola appearance and black spots around calcifications and MCCs, particularly in cancer cases (16%).
GAN General Limitations	Artifact incidence (69–78%) in normal tissue. Multiple artifacts may appear simultaneously in the same simulated mammogram (38%). This increases the risk of false positives during detection.
Data Requirements	Need for large training datasets. Risk of overfitting when data is limited.
Interpretability	Limited interpretability of GAN-generated outputs restricts clinical applicability.
Three-dimensional Model Limitations	Extensive parameter tuning to model realistically small structures due to voxel size constraints. These methods also require accurate characterization for each imaging system.
Hybrid Simulation Framework Limitations	Lesion integration with breast textures is challenging and may incorrectly focus on pectoral muscle or skin folds. Accurate breast masks are required, and generating correlated CC and MLO views remains difficult.
Validation and Assessment Limitations	Validation studies are commonly restricted to a limited number of radiologists and datasets. Multi-vendorvalidationisrare, establishingreliablegroundtruthischallenging, andverification using magnification images is often impractical. Validation outcomes may also depend on the observer.
General/Other	Some methods are currently limited to simulating single MCs instead of full clusters.

). Other challenges include the need for large training datasets, the risk of overfitting, and limited interpretability. These restrict clinical applicability.

Finally, the integration of DL tools into clinical practice remains complex. Challenges include regulatory constraints, interoperability with existing medical workflows, and the need for clinician training [13]. Table 11 summarizes these limitations.

7. Synthesis of Research Questions

Based on the evidence summarized in Section 6, and the synthesis of 72 studies, this review examined AI-driven MC detection in digital mammography, addressing three key research questions.

First, regarding the performance between DL and traditional methods, the 72 studies show that DL architectures, especially CNNs, have tended to report higher performance than traditional ML models. However, a key caveat should be noted: the 72 studies re- viewed used different datasets (e.g., MIAS, INbreast, DDSM, CBIS-DDSM, MEXBreast), different evaluation metrics (e.g., accuracy, sensitivity, AUC, F1-score), and different validation strategies (e.g., k-fold cross-validation, hold-out, external validation). Consequently, direct comparisons of reported performance values across studies are not methodologically robust. The observation that DL models have tended to report higher performance should be interpreted as a qualitative trend, not as evidence of universal superiority.

Table 10 summarizes the reported performance of MC detection for rule-based, ML-based, and DL-based methods. Performance metrics were reported across different studies and datasets. Traditional ML methods like RF and XGBoost remain effective in small-data scenarios and work well with structured clinical inputs. These methods provide stable performance. However, their maximum reported accuracy is 97.31% [62]. DL models have achieved reported accuracy as high as 99.71% [11]. Across the reviewed studies, DL methods have generally reported superior diagnostic performance for mammography-based MC detection, achieving the highest sensitivity and an AUC of 0.998 [95]. These gains, however, come with notable trade-offs, including heavy reliance on large datasets, high computational demands, and limited model interpretability.

Second, concerning the impact of explainable AI on clinical integration, existing studies suggest that XAI may enhance transparency, interpretability, and trust in AI systems [72]. Nevertheless, robust quantitative evidence demonstrating significant improvements in diagnostic confidence or seamless integration into clinical workflows remains limited [108].

Third, regarding the role of synthetic data in addressing dataset limitations, our analysis shows that synthetic data generation using DL models provides only partial mitigation of data size, diversity, and privacy constraints. Synthetic data can increase volume and reduce reliance on sensitive information. However, its effectiveness is constrained by data quality issues. Artifact incidence is high (up to 78% in normal samples). Multiple artifacts often co-occur in 38% of cases. Failures include black spots in 16% of cancer cases, breast boundary artifacts (34–53%), and checkerboard artifacts in 39–46% of simulated cases. These non-physiological patterns bias learning and reduce model robustness. Be- yond GAN-specific artifacts, broader limitations further constrain synthetic data utility [112]. Three-dimensional models require extensive parameter tuning and accurate system characterization. Hybrid simulation frameworks struggle with lesion integration and generating correlated CC/MLO views. Validation practices remain restricted to small datasets, with multi-vendor validation rarely performed and ground truth is difficult to establish.

8. Risk of Bias Due to Missing Results

For the synthesis of Research Question one (DL vs. traditional methods), the overall risk of bias was judged as moderate to high. Primary concerns included publication bias, which favors positive DL results, underreporting of null or negative findings, and inconsistent outcome reporting across studies.

For the synthesis of Research Question two (XAI on clinical integration), the risk of bias was assessed as high. The review identified a significant evidence gap, as many XAI methods (e.g., Grad-CAM saliency maps) do not necessarily reflect true model decision- making. These techniques often highlight visually salient pixels rather than causal features, producing explanations that are convincing but not always medically reliable, particularly for MCs. The literature also tends to emphasize positive qualitative claims, while evidence on clinical effectiveness, radiologists’ diagnostic confidence, and implementation barriers is still scarce.

For the synthesis of Research Question three (role of synthetic data in addressing dataset limitations), the risk of bias due to missing or incomplete evidence was assessed as moderate to high. While synthetic data generation has been proposed as a strategy to mitigate data scarcity and privacy constraints, the current body of evidence remains limited and heterogeneous. In particular, there is insufficient empirical validation regarding the extent to which synthetic data preserves clinical realism and supports robust generalization across diverse clinical settings.

9. Certainty of Evidence and Sensitivity Analyses

A formal application of the Grading of Recommendations, Assessment, Development and Evaluation (GRADE) approach for this review was not performed because the 72 studies we analyzed were highly heterogeneous in study design, measured outcomes, and reporting standards. Moreover, we did not conduct a meta-analysis.

For the comparison between DL and traditional methods (Synthesis 1), we rated the certainty of the evidence as moderate. DL models consistently outperformed traditional approaches across multiple studies and datasets. However, caution is warranted. The field lacks standardized reporting metrics, and there are very few prospective clinical validation studies. In addition, variations in evaluation protocols and database characteristics may affect direct comparisons between methods.

For the impact of explainable AI on clinical integration (Synthesis 2), we rated the certainty of the evidence as very low, as most of the evidence we found was qualitative or descriptive. These studies describe what XAI could or should do, but they do not measure its actual impact in real clinical settings. Few, if any, provide quantitative data on outcomes such as improvements in radiologists’ diagnostic confidence, reductions in unnecessary biopsies, or enhanced reading workflow efficiency. This represents a significant gap in the literature.

For the role of synthetic data in addressing dataset limitations (Synthesis 3), we rated the certainty as low. Synthetic data has shown promise, particularly in settings where real data is scarce. However, we are not yet convinced that synthetic data can fully capture the complexity of real clinical cases. Models trained on synthetic data often struggle to generalize. This limitation becomes evident when models are tested on different patient populations or imaging systems.

Taken together, these informal certainty ratings point to a clear need for more rigor- ous clinical validation. AI-driven MC detection must move from research labs into real clinical practice. These trials should be prospectively designed and standardized. Al- though the evidence base is promising, it is not yet solid enough to support widespread implementation.

A limitation of this review is the absence of sensitivity analyses to test the robustness of the synthesized findings. However, given the substantial heterogeneity across studies and the qualitative nature of the synthesis, sensitivity analyses were not deemed appropriate. Future updates of this review may consider including them if the evidence base becomes more homogeneous.

The present review was not prospectively registered in a systematic review registry (e.g., PROSPERO [113]), and no review protocol was prepared prior to its conduct. PROS-PERO [113] primarily accepts registrations for systematic reviews with meta-analysis. Our review follows a scoping review methodology with narrative synthesis. Therefore, the absence of registration is methodologically consistent. Nevertheless, we acknowledge this as a limitation. We encourage future updates to consider prospective registration where applicable.

10. Future Directions and Conclusions

The literature reviewed in this study highlights both the progress achieved in AI-driven MC detection and the challenges that continue to limit its clinical adoption. The following subsections discuss potential directions for future research and summarize the main conclusions derived from this review.

10.1. Future Directions

Future research should prioritize the development of explainable and clinically trust- worthy AI systems. Although DL models have achieved remarkable performance in MC detection, their widespread adoption in clinical practice depends on transparency and interpretability. The literature increasingly emphasizes the importance of model explain- ability to foster clinician trust and facilitate integration into diagnostic workflows [108]. Consequently, understanding and communicating the decision-making processes of DL models remain critical challenges [114].

Another important research direction is the development of large, diverse, and longitudinal mammography datasets. Emerging research aims to bridge the gap between single-time-point DL models and longitudinal screening analysis. CNN-based approaches can learn nonlinear subtraction filters that suppress normal anatomical variation while enhancing true pathological changes [115]. Beyond detection, DL models may characterize temporal changes in MCs, including the appearance of new lesions, cluster growth, morphological transformation, or stability, thereby supporting BI-RADS assessment of change categories [116]. Public resources such as OPTIMAM and EMBED represent important advances toward this goal [117,118]; however, larger, standardized, and publicly accessible longitudinal datasets are still needed.

Synthetic data generation has emerged as a promising solution for improving dataset diversity while mitigating privacy concerns [119]. Synthetic mammograms can enhance the representation of underrepresented populations and facilitate extensive model development and validation while preserving patient confidentiality [110]. Several studies have reported that models trained with synthetic data can achieve performance comparable to or exceeding that obtained with real data alone [120,121].

Future studies should also focus on improving model robustness and generalization. Ensemble and hybrid learning strategies have demonstrated potential for enhancing detection performance by leveraging complementary strengths across multiple models [80]. Similarly, the fusion of diverse DL architectures may improve robustness, reduce model variance, and increase generalizability across heterogeneous datasets and imaging conditions [122].

The responsible integration of AI into clinical practice will require continued attention to ethical and regulatory considerations. Privacy protection, data security, algorithmic bias, and transparency remain significant barriers to the acceptance of AI-based systems in oncologic imaging [123]. Addressing these challenges will be essential for the deployment of DL technologies for breast cancer screening programs.

Finally, emerging technologies such as edge AI and quantum computing may provide new opportunities for medical image analysis. Although still in their early stages, these technologies could enhance computational efficiency, support complex DL workflows, and facilitate real-time deployment of AI systems in clinical settings [124,125,126,127].

10.2. Conclusions

This review synthesizes AI-driven MC detection in mammography by integrating three complementary perspectives within a SPICE-derived framework: the comparative performance of DL and ML traditional methods, the role of XAI in clinical adoption, and the potential of synthetic data to address limitations in data availability and diversity.

The literature suggests a gradual shift from rule-based systems and traditional machine learning approaches toward DL architectures. While conventional methods are generally interpretable, they are often limited by handcrafted features, parameter sensitivity, and restricted generalizability. In the studies reviewed, DL models, particularly CNN-based and hybrid architectures, have generally reported higher performance in con- trolled research settings by learning spatial, textural, and contextual features directly from mammographic images. However, direct comparisons are complicated by heterogeneity in datasets, evaluation protocols, and reporting standards.

Despite promising advances, several challenges continue to limit the translation of DL systems into routine clinical practice. These include dataset heterogeneity, limited external validation, insufficient model explainability, potential algorithmic bias, and the lack of standardized evaluation protocols. The evidence for XAI’s clinical impact remains of very low certainty, as most studies provide qualitative rather than quantitative validation in real-world settings. Future research should prioritize robust and explainable models trained on diverse, multi-institutional datasets that better reflect real-world clinical variability, and should include prospective validation studies.

The findings also reveal a shift from early studies based on relatively small screen-film mammography datasets toward modern investigations using larger FFDM databases. Nevertheless, the availability of high-quality, standardized, and longitudinal datasets remains a critical requirement for further progress.

Finally, DL should be viewed as a decision-support technology that may complement—but not replace—radiologists, provided that its limitations are properly understood and addressed. Continued collaboration among clinicians, data scientists, and engineers will be essential to ensure the responsible development, validation, and deployment of AI systems. Rigorous evaluation, ethical deployment, and interdisciplinary collaboration may improve breast cancer detection and patient outcomes using DL-based approaches, although further real-world evidence is required.

Table 12. Characteristics of included studies (rule-based and classical ML).

Author, Year	Design	Images	Method	Dataset(s)	Key Findings/Metrics
Yu (2010) [47]	MCs detection	20	Rule + ML-based	MIAS	Sensitivity: 94% (FP: 1.0/img) or 90% (FP: 0.65/img). Model-based and statistical textural features improve detection.
Halkiotis (2007) [42]	MCCs detection	53	ANN + Morphology	MIAS	Sensitivity: 94.7%, FP: 0.27/img. Topographic features improve CAD.
Papadopoulos (2008) [37]	MCCs detection	26	Rule-based	MIAS + Private	AZ = 0.932. WT enhancement and LRM improve CAD.
Malar et al. (2012) [45]	MCs detection/classif.	400	Extreme Learning Machine	DoD BCRP	Accuracy: 94%. Wavelet features outperform other texture features.
Mohanalin (2014) [46]	MCs detection	247	WT + Thresholding	MIAS + UCSF	FROC: Sensitivity 96.5%, FP: 0.36/img. Entropy-based thresholding improves accuracy.
Ge (2007) [44]	MCCs detection	192	Rule + Classifier + CNN	Private	FFDM sensitivity: 70%, 80%, 90% (FP: 0.07, 0.16, 0.63/img). SFM FP: 0.15, 0.38, 2.02/img. FFDM outperforms SFM.
Ciecholewski (2017) [36]	MCs detect. + seg.	200	Rule-based	MIAS + DDSM	Similarity: 80.5%, Overlap: 75.7%, Extra: 19.8%. Runtime: 0.83 s/ROI. Fast automated method.
Alasadi (2017) [35]	MCs detection	66	Rule-based	MIAS	Sensitivity: 93.1%, FP: 10%. Image enhancement improves detection but variability remains.
Linguraru et al. (2006) [43]	MCs detection	83	Biologically inspired	DDSM	Sensitivity: 100%, FP: ∼2/img. Method is robust and reproducible.
Bajcsi (2021) [51]	Early breast cancer detection	322	k-means, GLRLM, PCA, GA, DT, RF	MIAS	GLRLM features (45°, 90°) best. PCA outperformed GA. RF achieved 100% training and 70% test accuracy.
Zhang (2012) [55]	MCCs detection	267	TWSVM + subspace learning	DDSM	Accuracy: 90.16%, Sensitivity: 94.42%, FPR: 8.32% ± 1.04, AUC: 0.9677.
Khehra (2016) [54]	MCCs classification	200	MLFFBP-ANN, SMO-SVM	Private	SMO-SVM best with 90.16% accuracy. Linear SVM effective for MCC classification.
Aziz (2024) [64]	Calcifications classification	30	RF, SVM, k-NN	Private	RF achieved ∼97% accuracy, outperforming k-NN. Further improvements needed for clinical-grade performance.
Miron (2022) [65]	Placental MCs detection	150	k-NN, SVM	Private (placenta)	k-NN: 84.01% accuracy; SVM: 76%. Results improve detectability of placental MCs in clinical data.
Menon (2024) [59]	MCs detection	2500	AlexNet, SVM, DT, k-NN, NB	CBIS-DDSM, BCDR	SVM best among classical ML models. AlexNet achieved superior overall performance.
Liang (2022) [68]	MCs classification	5476	XGBoost	Private	Accuracy: 90.24%, AUC: 0.89. XGBoost outperformed several ML models. Feature engineering was critical.
Diaz (2014) [58]	MCs classification	200	SVM	Digitized mammograms	Best model: Gaussian SVM. Sensitivity: 84.0% (glandular), 87.1% (dense), 88.7% (fatty). Overall sensitivity 85.9%.
Prinzi (2024) [61]	MCs detect. + classif.	758	ML + Radiomics	Private	Best model: XGBoost. AUC: 0.83 (healthy), 0.856 (benign), 0.876 (malignant). Features align with prior clinical findings.
Fanizzi (2020) [62]	Classification	260 ROIs	Rule-based	BCDR	AUC: 0.9816, Accuracy: 97.31%, Specificity: 100%. Strong discrimination with minimal features. Significant improvement (p ≤ 0.01).
Suhail (2018) [63]	MCCs classification	288 ROIs	DT + classifier	DDSM	Accuracy: 91%.
Stelzer (2020) [128]	MCs classification	235	ML + texture analysis	Private	Combined texture analysis and ML effectively distinguish benign vs. malignant MCs.

,

Table 13. Characteristics of included studies (rule-based and classical ML methods continued).

Author, Year	Design	Images	Method	Dataset(s)	Key Findings/Metrics
Oliver (2012) [129]	MCs and MCCs detection	602	Dictionary-based	MIAS + Private	Sensitivity: 80% at 1 FP cluster/image (ROC and FROC analysis).
Quintanilla (2010) [130]	MCs detection	100	Enhancement + k-means	Private	Accuracy: 97.72%, Sensitivity: 98%, Specificity: 99.67%, AUC: 0.9875.
Brahimetaj (2022) [131]	Early MCs detection	450	RF, SVM, MLP, AdaBoost	Private	Random Forest performed best. Accuracy 77.03%, sensitivity 60.46%, specificity 89.77%, F1 76.35%, AUC 0.801.
Mahmood (2021) [99]	MCs detection + diagnosis	322	ML + radiomics	MIAS	AUC 0.90, sensitivity 98%, accuracy 98%. Improves diagnostic efficiency.
Fadil (2020) [9]	MCs classification	966	DWT + Random Forest	Private	Sensitivity 93%, specificity 97%, accuracy 95%, AUC 0.92.

,

Table 14. Characteristics of included studies (DL methods).

Author, Year	Design	Images	Method	Dataset(s)	Key Findings/Metrics
Pesapane (2023) [69]	Detection + classification	1986	AlexNet, ResNet18/34	Private	AlexNet performed best. Detection: sensitivity 98%, specificity 89%, AUC 0.98. Classification: sensitivity 85%, AUC 0.94.
Kang (2021) [79]	MCs classification	1579	ResNet	Private	Accuracy 81.54%, specificity 91.41%, PPV 81.82%.
Luna (2025) [11]	MCCs classification	12,000 patches	CNN	INbreast	Accuracy 99.71% with compact model (29 k parameters).
Teoh (2024) [80]	MCs classification	1500	DL ensemble	CBIS-DDSM	Ensemble confidence: 0.9305 for MCs and 0.8859 for normal cases.
Honjo (2022) [83]	Image quality evaluation	136	DL enhancement	Private	Super-resolution mammograms significantly improved quality (PIQE, p < 0.001).
Wang (2017) [86]	MCCs detection	2000	Deep CNN	DDSM	AUC 0.971 (CNN classifier) vs. 0.944 (MC detector). Global features improve discrimination of clustered MCs.
Cai (2019) [87]	MCs diagnosis	990	CNN + SVM	Private	Precision 89.32%, sensitivity 86.89%. Deep features outperform handcrafted features.
Schönenberger (2021) [132]	BI-RADS classification	268	CNN	Private	CNNs effectively classify MCs according to BI-RADS, supporting standardized diagnosis.
Leong (2022) [100]	MCs detection	322	ResNet50	CBIS-DDSM	Accuracy: ResNet50 97.58%, ResNet34 97.35%, VGG16 96.97%, AlexNet 83.06%.
Gerbesi (2023) [89]	MCs detection + classification	1000	DeepMiCa	Private	Automated pipeline improves detection and classification, supporting clinical decision-making and reducing unnecessary biopsies.
Kumar (2022) [73]	MCs classification	1547	CNN	CBIS-DDSM	Best accuracy 94%, sensitivity 97%, AUC 0.96 using AdaDelta optimizer.
Yurdusev (2023) [76]	MCs detection + classification	500	DL	DDSM	Accuracy 97.67%, outperforming baseline and Faster R-CNN.
Zhang (2019) [90]	Multi-scale detection	450	DL	MIAS	Accuracy 97.16%. Improves small-target detection by 5–10%.
Kallenberg (2016) [133]	Feature learning	500	Convolutional sparse autoencoder	Private	Learns discriminative features without prior assumptions; supports risk scoring.
Ayyadurai (2024) [134]	MCs and mass detection	400	ML + DL	Private	Accuracy, precision, recall, and F1-score all approximately 99%.

,

Table 15. Characteristics of included studies (DL methods continued).

Author, Year	Design	Images	Method	Dataset(s)	Key Findings/Metrics
Hernández (2025) [10]	MCs detection + classification	1284	CNN + GLCM + Gabor	Multiple	Localization accuracy 73.0%, classification accuracy 73.0%. Moderate performance across metrics.
Rehman (2025) [96]	Mammogram classification	80	VGG16	PINIM, DDSM	PINIM accuracy 96%, DDSM accuracy 95%. High sensitivity and precision across datasets.
Abdulqader (2025) [111]	Image synthesis comparison	80	Pix2Pix, SPADE GAN, WGAN	ACDC, CHAOS	SPADE GAN achieved best results: PSNR 36 dB, SSIM > 0.97, Dice 0.94, FID < 0.01.
Luna (2025) [81]	MCCs detection (transfer learning)	620	CNN	MEXbreast	Accuracy: 98.32%, 99.27%, and 89.17% for 50, 70, and 100 μm resolutions.
Suzuki (2017) [135]	Overview	N/A	DL	Multiple	Overview of deep learning methods and their mathematical foundations in medical imaging.
Wang (2016) [98]	MCs/masses discrimination	1000	DL	Private	Accuracy: 87.3% for MCs, 61.3% for masses, 89.7% for combined MCs, 84.8% overall. Combined analysis improves performance.
Hsu (2025) [84]	MCs detection + segmentation	150	U-Net, V-Net	Private	U-Net outperforms V-Net. Accuracy up to 85% with preprocessing.
Oyelade (2022) [136]	Dataset augmentation	2781	GAN	MIAS	Synthetic ROIs improve dataset quality and enhance DL model performance for breast cancer detection.
Joseph (2024) [137]	Dataset balancing	322	cGAN	MIAS	Class imbalance negatively affects DL performance in MC detection and classification tasks.
Singla (2025) [92]	Synthetic data comparison	262,599	VQ-VAE, GAN	MIAS	High-quality synthetic data improves model performance and reliability in medical imaging.
Biju (2024) [110]	Image synthesis	60,000	GAN	CelebA, CIFAR-10	Inception Score: 9.69 (CelebA), 10.79 (CIFAR-10). FID: 7.91 (CelebA). Improved stability and image quality with improved loss function.
Shia (2024) [138]	MCs detection	11,303	YOLOv8	Private	Accuracy 84.2%, F1-score 0.82, mAP 70.9%, recall 79.6%. Strong localization performance.
Shia (2025) [95]	MCs classification	3674	EfficientNet, ResNet	Private (BI-RADS 1–2,5–6)	EfficientNet models significantly outperformed ResNet architectures (p < 0.05). Best model: EfficientNet-B3 with accuracy 86.9%, AUC 0.998, weighted F1-score 0.869. No significant differences among EfficientNet variants. EfficientNet-B0 achieved comparable performance with faster inference time. Five-fold cross-validation used.
Wenjie (2025) [93]	DCIS vs. IDC classification	294 cases	ResNet101	Private	AUC 97%, accuracy 93%, sensitivity 94%, specificity 92%. Combined model outperforms individual approaches.
Ayşe (2023) [76]	MCs detection + classification	500	R-CNN, YOLOv4	DDSM	Difference filters improve MC visibility and detection performance.
Liwen (2024) [122]	MCs classification	428	ResNet18, DenseNet121	Private	Outperforms standalone CNN models. Grad-CAM improves interpretability and localization.
Banteng (2026) [77]	MCs detection + segmentation	1500	YOLOv7	VinDr-Mammo + Private	Dice: 89.19 ± 0.26%, 88.76 ± 0.32%. Boundary loss improves segmentation accuracy and reduces manual intervention.
Ke (2024) [82]	MCCs segmentation	500	Swin Transformer	FFDM-DBT, CBIS-DDSM	DSC: 86.52% (DBT), 50.78% (FFDM). Strong cross-domain segmentation performance.
Elumalai (2026) [70]	MCs classification	N/A	ViT	CBIS-DDSM	Dual-branch Vision Transformer effectively integrates complementary feature representations for MCs classification. Achieved accuracy 96.80%, AUC 0.982. Demonstrates improved feature fusion capability for subtle MC patterns.

and

Table 16. Studies that explicitly used a computer-aided detection (CAD) system for MCs/MCCs.

Author, Year	Design	Images	Method	Dataset(s)	Key Findings/Metrics
Al-Qdah (2005) [38]	MCCs detection	25	CAD system	Private (3 demographic groups)	Accuracy: 87%, 88%, 90%. Wavelet-based detection is generally effective, but dense breast tissue remains challenging.
Arodź (2006) [39]	MCCs detection	50	db4 Wavelet Transform	Private	Detection improvement: 3.9 and 2.8 (radiologist assessment). Wavelet transform-based CAD enhances MC detection in small-field mammograms.
Destounis (2004) [40]	CAD evaluation (FN reduction)	318	CAD + Double reading	Screening population	False negatives reduced from 31% to 19%. CAD improves detection of missed cancers.
Karale et al. (2019) [29]	MCCs detection	180	Multistage CAD	DDSM + INbreast + PGIMER-IITKGP	Two-dimensional NEO method shows superior sensitivity and false-positive performance, especially on PGIMER-IITKGP dataset.
Songyang (2000) [139]	MCCs detection	20	CAD system	Nijmegen	Sensitivity 90%, false positives 0.5 per image. Further validation with larger datasets is required.
Hayat (2014) [56]	MCs detection	181	CAD system (k-NN, SVM, ANN)	MIAS	SVM achieved the best performance with 83% accuracy, improving diagnostic performance.
Mabrouk (2019) [140]	CAD framework	181	ML-based CAD	MIAS	Accuracy 96% (integrated features) and 97% (invariant moments with ANN).
Shiri (2022) [141]	MCs detection	815	DL-based CAD	Private	Accuracy 96.7%, sensitivity 96.7%, specificity 96.7%, AUC 0.988.
Loizidou (2020) [96]	MCs detection + classification	320	SVM-based CAD	Private	Accuracy improved from 91.42% to 99.55% using temporal subtraction (p < 0.005).
Rehman (2021) [74]	MCs detection	6453	FC-DSCNN CAD	DDSM, PINUM	Detection score up to 0.97. True positive rate 0.99 (2.45 false positives per image). Strong performance but sensitive to preprocessing quality.
Savelli (2020) [60]	Thesis	500	DL-based CAD	Private	PhD thesis on deep learning-based CAD systems.
Sarvestani (2023) [85]	Image enhancement	250	ML-based CAD	DDSM	Sensitivity 95.7%, specificity 91.5%, precision 89.3%, accuracy 93%.

summarize the included studies and support the comparative analysis presented in this review.