Answers are grounded in this paper and cite their sources.
Digital mammography is increasingly analyzed with radiomics and deep learning, yet reproducibility remains constrained by heterogeneity in acquisition systems, Digital Imaging and Communications in Medicine (DICOM) presentation semantics, and preprocessing pipelines. This narrative review, based on a documented Scopus search that screened 5503 records, synthesizes computational strategies for harmonizing full-field digital mammography before quantitative analysis. Evidence is organized into four levels (representation-aware preprocessing, image-domain, feature-domain, and model-level harmonization), and mammography-specific studies are distinguished from evidence obtained in computed tomography and magnetic resonance imaging. The literature indicates that stored pixel values, photometric interpretation, and modality and value-of-interest (VOI) transformations must be resolved before radiomic or learned representations are compared; that feature reproducibility requires dedicated study designs rather than between-group comparisons; and that feature-level corrections such as ComBat require training-only fitting and protection of biological covariates. A descriptive audit of 1622 VinDr-Mammo DICOM images illustrates how representation modes co-vary with the equipment manufacturer and specifies, with released code, a protocol for comparing alternative preprocessing pipelines on identical images. The review identifies gaps in DICOM reporting, vendor-stratified validation, patient-level leakage control, and external evaluation, and maps each step of a proposed workflow to its level of evidence.
No takes yet. Share an insight, caveat, or question.
Orazayeva et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: