OPTIMAM: Mammography Research Database
- OPTIMAM is a large-scale, curated mammography research programme combining millions of images with screening, clinical, and pathological data.
- It supports diverse studies including computer-aided detection, risk prediction, virtual clinical trials, and cross-site generalization through rich lesion annotations and metadata.
- Its infrastructure features automated and stand-alone data collection with pseudonymisation and controlled access, ensuring high data integrity and ethical research governance.
OPTIMAM is a Cancer Research UK–funded mammography research programme, originally created to support OPTIMAM (2008–2013) and OPTIMAM2 (2013–2018), whose central data asset is the OPTIMAM Mammography Image Database (OMI-DB): a large, curated, shareable research database of mammography images linked to screening, clinical, and pathological information (Halling-Brown et al., 2020). OMI-DB consists of several relational databases and cloud storage systems and, in the main content description, contains 2,889,312 images of all types from 173,319 women; the abstract summarizes this as over 2.5 million images from the same cohort (Halling-Brown et al., 2020). Its combination of mammographic imaging, follow-up, interval cancers, prior examinations, and lesion-level expert marks has made OPTIMAM a recurrent source domain for computer-aided detection, risk prediction, image perception studies, virtual clinical trials, and cross-site generalization research (Morrell et al., 2019).
1. Programme origins and population scale
OPTIMAM was built to address a structural limitation in medical-imaging research: the lack of very large, well-curated, sharable datasets with trustworthy outcome labels and longitudinal follow-up (Halling-Brown et al., 2020). Within the programme, OMI-DB was designed not only for the original OPTIMAM studies but also for broader use in virtual clinical trials (VCT), computer-aided detection (CAD), artificial intelligence / machine learning, image perception studies, reader training, and quality assurance (Halling-Brown et al., 2020). The database was collected from three UK breast screening centres: Jarvis Breast Screening Centre, Guildford; St George’s Hospital, South West London; and Addenbrooke’s Hospital, Cambridge (Halling-Brown et al., 2020).
The cohort reported in the abstract is stratified as follows (Halling-Brown et al., 2020):
| Group | Women |
|---|---|
| Normal breasts | 154,832 |
| Benign findings | 6,909 |
| Screen-detected cancers | 9,690 |
| Interval cancers | 1,888 |
The collection strategy combined exhaustive ascertainment and sampling. It included all screen-detected cancers from the three sites, prior screening mammograms of interval cancers since 2012, all women screened between 1 January 2014 and 31 December 2014, and a random 25% sample of all women screened in 2012, 2013 and 2015 (Halling-Brown et al., 2020). Collection is ongoing, and women are followed up with their clinical status updated according to subsequent screening episodes (Halling-Brown et al., 2020). This longitudinal design makes OPTIMAM more than a cancer archive; it is a screening-population resource containing normal, benign, screen-detected, and interval-cancer trajectories.
2. Database structure, imaging content, and lesion annotation
OMI-DB stores both images and linked metadata extracted largely from the National Breast Screening System (NBSS), including radiological information, clinical information, and pathological information (Halling-Brown et al., 2020). The imaging component includes screening mammograms, both processed and unprocessed images, initially 2D digital mammography, and later additional modalities such as tomosynthesis and MRI (Halling-Brown et al., 2020). When images are loaded, relevant DICOM tags are extracted into a searchable index; fields shown in the schema include study_id, series_id, modality, manufacturer, view_position, image_laterality, and presentation_intent_type (Halling-Brown et al., 2020).
The relational schema links the hierarchy from woman to image. The simplified schema includes entities such as client, episode, episode_event, imaging, study / series / images, lesion, marks, and clinical/surgery/biopsy data (Halling-Brown et al., 2020). Representative fields shown in the paper include ClientID, EpisodeID, EventID, LesionID, Episode Type, EpisodeAction, EpisodeOpenedDate, EpisodeClosedDate, AssessmentType, DatePerformed, FinalAction, and MammoOpinion, together with lesion-side, lesion-position, and pathology-related fields such as in-situ grade and invasive grade (Halling-Brown et al., 2020). This structure supports linkage across screening history, biopsy results, surgery, and image-level annotations.
A major value-added feature is lesion localization. Because lesion location is not routinely stored in clinical databases, it was collected specifically for OMI-DB by experienced mammography readers, who manually annotated lesion location, lesion area, radiological appearance, and conspicuity (Halling-Brown et al., 2020). The paper reports that 7,143 lesions have been marked, including approximately 60% of screen-detected cancers (Halling-Brown et al., 2020). The schema’s marks table includes fields such as mark_id, ImageID, mark_type, mark_shape, radiological_appearance, and conspicuity (Halling-Brown et al., 2020). This makes OPTIMAM suitable not only for whole-exam classification but also for lesion localization and detection studies.
3. Ingestion workflow, de-identification, and controlled access
OMI-DB is architected as several relational databases and cloud storage systems, with distributed collection and centralized storage for research use (Halling-Brown et al., 2020). Two collection modes are described: automated remote-site collection and stand-alone collection (Halling-Brown et al., 2020). In the automated mode, a physical or virtual server on site queries NBSS for women matching collection requirements such as date range, outcome classification, or high-risk status, and then retrieves images and clinical data for screening and assessment episodes (Halling-Brown et al., 2020). In the stand-alone mode, a tool can ingest images from a prepared folder while querying NBSS for the relevant clients (Halling-Brown et al., 2020).
At collection time, imaging and screening data are pseudonymised, lookup tables are created at the collection site, and images, image metadata, and screening data are uploaded to the cloud for storage in the central database (Halling-Brown et al., 2020). For external sharing, the resource applies additional de-identification, and a dedicated database records which cases were shared with each third party, together with investigator and sharing metadata (Halling-Brown et al., 2020). All collection activity is logged, including NBSS queries, PACS queries, metadata on collection runs, and client-related collection activity (Halling-Brown et al., 2020). A web-enabled application, MedxViewer, supports remote case viewing, feature annotation, and observer studies (Halling-Brown et al., 2020).
Governance is integral to the OPTIMAM model. The project has approval from an ethical research committee specialising in research databases, organised by the NHS Health Research Agency (Halling-Brown et al., 2020). Cancer Research UK (CRUK) retains the intellectual property of the database, and sharing occurs through agreements with approved academic and commercial research groups (Halling-Brown et al., 2020). The database has been shared with over 30 research groups and companies since 2014 (Halling-Brown et al., 2020). An open-source Python package is also mentioned for parsing shared OMI-DB data and providing an API, metadata extraction, and filtering support (Halling-Brown et al., 2020). Accordingly, OPTIMAM is a controlled-access research infrastructure rather than an unrestricted public-download dataset.
4. OPTIMAM as a lesion-supervised source domain for mammography CAD
In later CAD work, OPTIMAM functioned as a large lesion-supervised training source rather than merely an exam-level label repository. A high-resolution mammography CAD study based on R-FCN / DCN used a recently released OPTIMAM dataset consisting of about 78,000 selected digital screening and symptomatic mammograms, containing approximately 7,500 findings annotated with bounding boxes, about 6,500 of them cancerous, with biopsy-determined ground truth and images mainly from Hologic and GE machines at around pixels (Morrell et al., 2019). The authors state that they selected a detection architecture even though the final task was whole-image classification because they wanted to “exploit the rich bounding box information in our training dataset (Optimam)” and improve interpretability (Morrell et al., 2019).
That study trained on all OPTIMAM images with findings and used a three-class detector with classes negative, benign, and malignant findings, using the malignant class score for downstream malignancy prediction (Morrell et al., 2019). The preprocessing explicitly described for OPTIMAM was manufacturer-specific intensity normalization via lookup tables, reflecting the fact that the data span multiple acquisition vendors (Morrell et al., 2019). The architectural choices were tightly coupled to OPTIMAM’s image scale and lesion granularity: the model used an input size of pixels, batch size one image per GPU, and mammography-specific rotation/flip augmentation while avoiding random crops because lesions near image borders could be lost (Morrell et al., 2019).
The paper did not report a dedicated OPTIMAM train/validation/test split or a final held-out OPTIMAM benchmark. Instead, OPTIMAM served as the strongly annotated source domain that enabled transfer to the independent DREAMS / Group Health benchmark, where the final detector achieved AUC for breast-wise detection on 130,000 hidden validation images (Morrell et al., 2019). This use of OPTIMAM established a pattern that recurs in subsequent work: lesion-level supervision in OPTIMAM is exploited to train models whose principal quantitative evaluation occurs on separate cohorts.
5. OPTIMAM in short-horizon risk prediction and cross-site domain generalization
OPTIMAM has also been used as a multi-site population resource for fixed-horizon risk prediction. In MamaDino, the source population came from the OPTIMAM Mammography Image Database, specifically women attending routine FFDM screening at four UK screening services—Jarvis, Leicester, Imperial, and St George’s—between 2010 and 2021; after filtering, the training cohort comprised 53,883 women (Santeramo et al., 14 Feb 2026). Labels were built from episode outcomes and lesion laterality: episodes with CI / M / B outcomes and their screening exams in the preceding 3 years were labeled positive, normal episodes were labeled negative, and lesion-side OPTIMAM annotations supplied breast-level supervision (Santeramo et al., 14 Feb 2026). Using 512×512 mammograms, the hybrid MamaDino + BilateralMixer model reached AUC 0.736 on an internal matched case-control test set and AUC 0.677 on an external out-of-distribution cohort from an unseen OPTIMAM site, Oxford, while the paper states that this matched Mirai despite using ~13× fewer input pixels (Santeramo et al., 14 Feb 2026). A plausible implication is that OPTIMAM’s bilateral structure and lesion laterality enable both breast-level and patient-level modeling strategies that would be difficult to construct from image-only repositories.
A different line of work used OPTIMAM as the sole supervised source domain for malignant-versus-benign calcification classification across institutions. That study used 2,994 annotated calcification cases from OPTIMAM, including 2,187 malignant cases (73.0%) and 807 benign cases (27.0%), and performed 5-fold cross-validation using only OPTIMAM during training and validation (Liu et al., 7 Jul 2026). The labeled subset was effectively dominated by Hologic FFDM screening data, while additional unlabeled OPTIMAM subdomains—GE FFDM and Hologic synthetic images—were mined for domain adaptation (Liu et al., 7 Jul 2026). The baseline Swin Transformer V2 achieved AUC on OPTIMAM, but external performance fell to 0.68 on both EMBED and Duke; adding the paper’s unsupervised style-transfer adaptation improved AUC from 0.68 to 0.72 on EMBED and from 0.68 to 0.73 on Duke (Liu et al., 7 Jul 2026). This use case shows that OPTIMAM is large and heterogeneous enough to support both supervised lesion classification and adaptation experiments, while also revealing that source-domain strength does not eliminate cross-site domain shift.
6. Representativeness, limitations, and scientific position
OPTIMAM’s scale and linkage are accompanied by real-world limitations that matter for experimental design and interpretation. The 2020 systems paper states that lesion locations are not routinely present in clinical systems, so only a subset have manual marks; 7,143 lesions represent only about 60% of screen-detected cancers (Halling-Brown et al., 2020). Some pathology fields are incomplete in the source systems: 20 screen-detected cancers and 12 interval cancers were excluded from the grade table because invasive-status and grade information were missing from the clinical databases (Halling-Brown et al., 2020). The cohort assembly includes both exhaustive and sampled components, so years and classes are not uniformly sampled (Halling-Brown et al., 2020). The resource reflects the UK NHS Breast Screening Programme, where women are typically screened every three years between ages 50 and 70, with some additional age-trial and high-risk screening groups (Halling-Brown et al., 2020). This suggests that OPTIMAM is highly relevant to UK population-screening research but should not be assumed to transfer unchanged to settings with different intervals, age ranges, or case mix.
Within those constraints, OPTIMAM is distinctive because it combines large scale, linked imaging + clinical + pathology data, processed and unprocessed images, prior mammograms, interval cancers, partial lesion localization, longitudinal updating, and a governance model that has already supported substantial academic and commercial reuse (Halling-Brown et al., 2020). Internal uses reported in the database paper include virtual clinical trials, studies of detector type, dose, and image processing on cancer detection, analyses of cancer characteristics, and studies of breast density in NHSBSP women (Halling-Brown et al., 2020). Across later work, OPTIMAM has served as a lesion-annotated detector-training set, a bilateral risk-prediction cohort, and a source domain for multi-site adaptation (Morrell et al., 2019). In that sense, OPTIMAM is best understood not as a single benchmark split but as a reusable mammography research infrastructure whose scientific value derives from the joint availability of images, metadata, pathology, follow-up, and partially localized lesions.