https://atlas.rsna.org/schemas/2025-11/model.json

Deep Learning for Breast Cancer Risk Prediction: Application to a Large Representative UK Screening Cohort

https://dx.doi.org/10.1148/ryai.230431

1.0

ten.shn@2sille.mas

Creation and maintenance of the OPTIMAM Image Database funded by Cancer Research UK (C30682/A28396). S.E. supported by the Million Women Study (C16077/A29186).

Data were collected with approval from an ethical research committee specializing in research databases organized by the NHS Health Research Authority.

Pretrained ShuffleNet V2 backbone followed by fully connected layers; trained with focal binary cross-entropy loss and balanced batches; data augmentation with random vertical flips and rotations.

Predicts 3-year future breast cancer risk from current negative screening mammograms to enable potential risk-stratified screening and earlier detection, particularly of interval cancers.

Clinical decision support systems; potential workflow optimization for risk-stratified screening.

Operating points reported include balanced, high-sensitivity (95%), and high-specificity (95%) settings; specific score thresholds not reported.

Automated risk scoring to support, not replace, clinician decision-making.

Women aged 50–70 years attending routine triennial UK NHS Breast Screening Programme; prediction of future (within 39 months) breast cancer risk from a current negative screening examination.

Single mammographic view (e.g., CC or MLO “for presentation” DICOM) and patient age.

Apply model to ‘for presentation’ screening mammograms (two views per breast typically available) acquired on Hologic systems; compute patient-level risk as the mean of image-level scores; include patient age as an input.

Evaluated on three UK sites using Hologic systems only; generalizability to other manufacturers and independent sites not yet established. Single-view input with post hoc patient-level aggregation; does not leverage longitudinal prior exams or non-imaging data beyond age. Excludes images with implants. Further clinical pathway design and cost–benefit analyses required; not validated as a diagnostic tool.

Further external validation across additional sites and vendors and integration studies within NHSBSP are recommended before clinical deployment.

Model selection based on highest validation AUC; evaluation on a hold-out test set; analyses performed in Python 3.8 using NumPy, pandas, SciPy, and scikit-learn. Full training details and hyperparameters in supplemental materials.