Recall Rate

High fibroglandular tissue density creates a radiographically dense background that can obscure underlying malignancies. In patients with ACR Category C or D density, lesions and normal parenchyma appear similarly white, reducing anatomical contrast. This masking effect directly elevates recall rates as radiologists and AI systems flag ambiguous overlapping structures. Automated density classification algorithms can normalize detection thresholds based on the density category to mitigate this variability.

The operating point on the Free-Response Operating Characteristic (FROC) curve is a tunable parameter that directly governs recall behavior. Lowering the detection threshold increases true positive sensitivity but generates more false positive marks per image. This trade-off is often adjusted based on clinical context:

• Screening workflows prioritize high sensitivity at the cost of higher recalls
• Diagnostic workflows may tolerate lower sensitivity for higher specificity
• Model calibration ensures predicted confidence scores align with empirical malignancy rates

The transition from Full-Field Digital Mammography (FFDM) to Digital Breast Tomosynthesis (DBT) has systematically reduced recall rates by 15-30%. DBT's quasi-3D reconstruction resolves tissue superposition artifacts that mimic architectural distortion in 2D projections. However, the increased data volume per exam introduces new challenges:

• Reading time increases without AI-assisted triage
• Synthetic 2D images generated from DBT data reduce dose but may alter texture features
• Maximum Intensity Projection (MIP) slabs can accelerate calcification review

Access to prior exam registration enables the detection of subtle interval changes that are invisible in a single study. Deformable registration algorithms spatially align current and historical mammograms, allowing subtraction techniques to highlight new or evolving findings. The absence of priors—common in baseline screenings or when patients switch providers—forces a more conservative reading strategy, significantly increasing recall rates due to the inability to confirm stability of benign-appearing asymmetries.

The clinical integration paradigm—concurrent reading versus second-reader CADe—fundamentally alters recall behavior. In concurrent workflows, AI marks are displayed during initial interpretation, which can anchor the radiologist's attention but may also introduce automation bias. Key interaction factors include:

• Mark saliency and boundary cues influence perceptual attention
• False positive reduction post-processing suppresses erroneous marks
• Worklist prioritization triages high-suspicion exams for immediate review
• Inter-reader variability in how different radiologists incorporate or override AI suggestions

The inherent visual characteristics of a finding directly impact detection difficulty and recall probability. Spiculated masses with radiating lines are highly conspicuous and rarely missed, while architectural distortion—a focal retraction without a central mass—represents one of the most challenging findings. Microcalcification clusters, particularly amorphous or punctate morphologies, require high-resolution patch-based analysis to distinguish benign from suspicious distributions. The subtlety gradient correlates strongly with false negative rates and the decision threshold for recall.

The proportion of actual cancers correctly identified by the screening system. Sensitivity and recall rate are mathematically identical but used in different contexts—sensitivity describes the test's ability to find disease, while recall rate describes the operational burden on the health system. A system with 95% sensitivity correctly flags 95 of 100 cancers.

• Calculated as: TP / (TP + FN)
• High sensitivity is non-negotiable in screening mammography
• The clinical goal is maximizing sensitivity while minimizing unnecessary recalls

The proportion of healthy patients correctly cleared by the screening exam. Specificity is the direct counterbalance to recall rate—every false positive recall reduces specificity. A screening program with 90% specificity correctly clears 90 of 100 cancer-free women.

• Calculated as: TN / (TN + FP)
• Low specificity drives unnecessary diagnostic workups
• AI computer-aided detection systems aim to improve specificity without sacrificing sensitivity

The probability that a recalled finding actually represents cancer. PPV directly measures the efficiency of the recall process—a low PPV indicates excessive false alarms. In screening mammography, PPV varies significantly by patient age, breast density, and the specific finding type.

• PPV1: Cancer yield from abnormal screening exams
• PPV2: Cancer yield from biopsies recommended
• PPV3: Cancer yield from biopsies performed
• A recall rate of 10% with a PPV1 of 5% means only 1 in 20 recalls yields cancer

A statistical analysis curve that plots the true positive detection rate against the average number of false positives per image, used to evaluate localization performance. Unlike ROC analysis, FROC accounts for the spatial dimension—a detection is only counted if the algorithm correctly localizes the lesion.

• X-axis: False positives per image (non-cumulative)
• Y-axis: Lesion-level sensitivity
• Essential for comparing CADe systems at clinically relevant operating points
• A lower FROC curve indicates more false marks at equivalent sensitivity

The Breast Imaging Reporting and Data System standardizes recall decisions. BI-RADS 0 indicates an incomplete assessment requiring additional imaging—this is the category that directly drives recall rate. Understanding the distribution of BI-RADS assignments is essential for interpreting recall metrics.

• BI-RADS 0: Incomplete—need additional imaging (triggers recall)
• BI-RADS 1: Negative (routine screening)
• BI-RADS 2: Benign finding (routine screening)
• BI-RADS 3: Probably benign (short-interval follow-up)
• BI-RADS 4: Suspicious abnormality (biopsy recommended)
• BI-RADS 5: Highly suggestive of malignancy