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

Semiautonomous Deep Learning Breast Cancer Rule-out System for Screening Mammography

https://dx.doi.org/10.1148/ryai.230033

1.0

ia.tibbaretihw@hcraeser

Supported by Whiterabbit.ai. Washington University in St. Louis has equity interests in Whiterabbit.ai and may receive royalty income and milestone payments per an agreement with Whiterabbit.ai to develop the technology.

IRB approvals obtained; informed consent waived; data handled according to HIPAA.

Two-stage deep learning system: low-level vision model analyzes each mammogram image independently; high-level vision metamodel integrates multi-view, bilateral, and prior imaging along with age and prior BI-RADS assessments to produce an exam-level malignancy probability.

Reduces screening mammograms requiring radiologist interpretation, decreases false-positive callbacks and benign biopsies while maintaining cancer detection rate within a predefined noninferiority margin.

Patients’ triage; workflow optimization; clinical decision support for screening mammography.

Rule-out operating threshold selected on validation data to achieve target 12-month sensitivity of 99% (and alternatively 97%). Exams with scores below threshold are assigned BI-RADS 1 by the device.

Semiautonomous rule-out: automatically assigns BI-RADS 1 to a subset of nonsuspicious screening exams; remaining exams proceed to standard radiologist interpretation.

Identification of screening mammograms not suspicious for breast cancer in a screening population, to reduce radiologist workload and false-positive downstream events in screening programs (US and UK settings).

2D full-field digital mammography images (multi-view, bilateral), patient age, prior mammogram images, and prior exam BI-RADS assessments when available.

Device reads examinations before radiologists. Exams with prediction below the rule-out threshold are automatically labeled BI-RADS 1 (negative) and removed from radiologist workflow; all others are read by radiologists as usual. Site- and scanner-specific operating point selection may be considered.

Retrospective simulation; assumes radiologist behavior unaffected by removal of exams. Performance varies by scanner model (HSE vs SED) due to training data imbalance. Limited tracking of false negatives in one dataset (US2). Test sets include multiple exams per patient which may underestimate variance. Radiologist false negatives include interval cancers likely not visible at screening. Prospective reader studies and QA/monitoring systems are needed.

Operate at high-sensitivity rule-out thresholds; consider site/scanner model when selecting operating point; further prospective validation and QA monitoring recommended.

Evaluation conducted on retrospective datasets with specified splits; performance reported with prevalence adjustments, noninferiority tests, and bootstrap CIs. Code not provided.