Screening Mammography

Screening mammography is performed on a defined, asymptomatic population based on age and risk factors, not clinical symptoms. The goal is to detect preclinical disease before it becomes palpable or symptomatic. This distinguishes it from diagnostic mammography, which investigates specific clinical findings. AI models trained on screening populations must be optimized for low disease prevalence (typically 3–7 cancers per 1,000 exams), where maintaining high specificity is critical to avoid overwhelming recall rates.

A standard screening exam consists of two projection views per breast:

• Craniocaudal (CC): A top-to-bottom compression view
• Mediolateral Oblique (MLO): An angled side view capturing axillary tissue

This dual-view protocol ensures maximum tissue visualization and provides geometric redundancy. AI detection systems must perform multi-view correlation to confirm that a suspicious finding appears in both projections, significantly reducing false positives caused by superimposed tissue shadows.

Screening mammograms are typically interpreted in high-volume batch reading sessions, where a radiologist reviews 50–100+ cases in a single sitting. This workflow creates fatigue-related observational errors as a primary failure mode. AI systems address this through worklist prioritization, reordering the queue so that suspicious exams are read first while the radiologist is freshest, and through persistent attention mechanisms that do not degrade over time.

The recall rate—the percentage of screened women called back for additional imaging—is a critical operational metric tightly regulated by accrediting bodies like the ACR and MQSA. Acceptable ranges typically fall between 5–12%. An AI system that is overly sensitive will drive recall rates above acceptable thresholds, causing unnecessary patient anxiety and biopsy procedures. Effective AI must balance cancer detection rate (CDR) against recall rate to demonstrate clinical utility.

Radiologists routinely compare a current mammogram against prior exams to identify subtle interval changes—new or enlarging masses, developing asymmetries, or new calcifications. This temporal context is often the only clue to an early malignancy. Advanced AI systems incorporate prior exam registration and temporal subtraction algorithms to automatically highlight changes over time, mimicking and augmenting the radiologist's comparative reading process.

All screening mammograms are interpreted and reported using the Breast Imaging Reporting and Data System (BI-RADS), a standardized lexicon that assigns a final assessment category from 0 to 6:

• BI-RADS 0: Incomplete, needs additional imaging
• BI-RADS 1: Negative
• BI-RADS 2: Benign finding
• BI-RADS 3: Probably benign (short-interval follow-up recommended)

AI systems must map their outputs to this established framework for seamless clinical integration.