From Machine Learning to Ensemble Approaches: A Systematic Review of Mammogram Classification Methods

Background/Objectives: Breast cancer remains one of the leading causes of mortality among women, necessitating continued advancements in diagnostic methods to enhance early detection and treatment outcomes. This review explores the current landscape of breast cancer classification, focusing on machine learning (ML), deep learning (DL), and hybrid/ensemble models. Methods: A systematic search following PRISMA guidelines identified 50 eligible studies published between 2018 and 2025. Studies were included based on their use of mammogram datasets and implementation of computer-aided diagnosis methods for classification. Models were compared in terms of preprocessing, feature extraction, optimization strategies, and classification performance. Results: Representative high performing models illustrate the strengths and limitations of each approach. In ML, an optimized ELM achieved 100% accuracy on MIAS. DL methods such as Vision Transformers also reached 100% accuracy on DDSM, outperforming conventional CNNs. Hybrid models, particularly IEUNet++, achieved 99.87% accuracy, offering robust multi class classification. Conclusions: While ML and DL approaches can achieve near perfect accuracy, they typically focus on binary classification tasks and require extensive preprocessing, feature extraction, and optimization. In contrast, hybrid methods provide comparable or superior performance while simultaneously addressing multi-classification with fewer handcrafted steps, highlighting their robustness. These findings underscore the need for innovative solutions that balance model accuracy, interpretability, and resource efficiency. By addressing these challenges, future classification systems can better support early breast cancer detection and improve patient outcomes.