Artificial Intelligence Pipeline for Mammography-Based Breast Cancer Detection: An Integrated Systematic Review and Large-Scale Experimental Validation

Resumen

Background and Objectives: Breast cancer remains a leading cause of cancer-related morbidity and mortality worldwide, and robust, interpretable artificial intelligence (AI) pipelines are increasingly being explored to support mammography-based detection. This study combines a PRISMA 2020-compliant systematic review with an original experimental validation to characterize current evidence and address identified gaps in reproducibility and interpretability. Materials and Methods: A PRISMA 2020-guided systematic review and an original experimental study were conducted. The review searched PubMed and Scopus/ScienceDirect for studies using convolutional neural networks (CNNs), support vector machines (SVMs) or eXtreme Gradient Boosting (XGBoost) for breast cancer detection in mammography and related imaging modalities, and identified 45 eligible articles. In parallel, we implemented and evaluated representative CNN (ResNet-50, EfficientNetB0 and MobileNetV3-Small) and classical machine learning (SVM and XGBoost) pipelines on the CBIS-DDSM dataset, following a CRISP-DM-inspired workflow and using Grad-CAM and SHAP to provide image- and feature-level explanations within a reproducible machine-learning-operations (MLOps)-oriented framework. Results: The systematic review revealed substantial heterogeneity in datasets, preprocessing pipelines, and validation strategies, with a predominant reliance on internal validation and limited use of explainable AI methods. In our experimental evaluation, ResNet-50 achieved the best performance (AUC-ROC 0.95; sensitivity 89%), followed by XGBoost (AUC-ROC 0.90; sensitivity 74%) and SVM (AUC-ROC 0.84; sensitivity 66%), while EfficientNetB0 and MobileNetV3-Small showed lower discrimination. Grad-CAM produced qualitatively plausible heatmaps centered on annotated lesions, and SHAP analyses indicated that simple global image-intensity and size descriptors dominated the predictions of the classical models. Conclusions: By integrating systematic evidence and large-scale experiments on CBIS-DDSM, this study highlights both the potential and the limitations of current AI pipelines for mammography-based breast cancer detection, underscoring the need for more standardized preprocessing, rigorous external validation, and routine use of explainable AI before clinical deployment.