Image clustering for breast cancer detection and diagnosis: a systematic review of algorithms, imaging modalities, feature representations and reported performance (2020–2025)

Background Breast cancer is the most common cancer in women globally, and many women die from it. Supervised deep learning is currently the dominant method for breast image analysis, but it relies on large annotated datasets that are expensive to collect and scarce in countries where breast cancer incidence is rising fastest. Although no pixel-level labels are needed for clustering, no previous synthesis has documented its applications across breast imaging modalities in the deep learning era. This review outlines the algorithms, modalities, feature representations and assessment metrics used in breast image clustering studies from 2020 to 2025. Methods After PRISMA 2020, we searched IEEE Xplore, PubMed/MEDLINE and Scopus for peer-reviewed, primary studies published in English between January 2020 and December 2025 in which image clustering was applied directly to breast images, with enough detail to appraise method, dataset and metrics. A single reviewer performed screening and deduplication in Rayyan, testing each candidate citation against its source record in the published version. The QUADAS-2, adapted for unsupervised segmentation, was used to assess risk of bias. Findings were synthesised narratively as study counts, given the heterogeneity of the included studies. Results Of 1552 records identified, 11 were eligible. Mammography was the most studied modality (six studies), followed by histopathology and dynamic contrast-enhanced MRI (two each) and ultrasound (one). Five studies used fuzzy C-means or a variant, four used K-means or a variant, and two used other partitioning methods; five clustered using a metaheuristic optimiser, and three clustered in a learned feature space. Supervised metrics (accuracy, sensitivity and specificity) dominated reporting; two studies reported per-image computational cost and two reported internal cluster-validity indices. No studies reported silhouette coefficients, inter-observer agreement or prospective validation. Conclusions The field has evolved from centroid-based partitioning to optimisation-enhanced and deep-feature clustering, but evaluation is still measured predominantly by agreement with supervised labels, and the properties of unsupervised methods remain poorly assessed. The main barriers to clinical translation are uneven reporting, reliance on a small number of publicly available datasets, and the absence of prospective, demographically diverse validation. This review has not been registered.