Interval Cancer

Interval cancers are disproportionately high-grade, rapidly proliferating tumors with elevated Ki-67 indices. They often lack the classic radiographic hallmarks of slow-growing malignancies.

• Frequently triple-negative or HER2-enriched molecular subtypes
• Higher likelihood of lymph node metastasis at diagnosis
• Shorter volume doubling times compared to screen-detected cancers
• Often present as architectural distortion rather than discrete masses

Dense breast tissue is the dominant factor in missed cancers. Fibroglandular tissue appears white on mammography—the same radiographic density as malignancies—creating a camouflage effect.

• ACR Density Category C (heterogeneously dense) and D (extremely dense) carry highest risk
• Masking accounts for 30-50% of interval cancers in dense breasts
• Digital Breast Tomosynthesis (DBT) reduces but does not eliminate masking
• AI models trained on dense breast cohorts show improved detection in this subgroup

A subset of interval cancers are visible in retrospect on the prior screening mammogram but were not flagged—a perceptual miss rather than a technical limitation.

• Retrospective review identifies visible findings in 20-40% of interval cancers
• Common missed signs: subtle asymmetries, developing densities, and one-view findings
• Satisfaction of search—stopping after finding one lesion—contributes to misses
• CADe systems reduce perceptual errors by providing a second observer that never fatigues

Interval cancers often arise from tissue that appeared completely normal on the prior exam, representing true de novo tumorigenesis rather than missed lesions.

• True interval cancers: no visible correlate on prior imaging even in retrospect
• Minimal sign cancers: subtle, non-specific findings visible only on review
• False negative cancers: clear findings missed at interpretation
• Temporal comparison algorithms that perform automated image registration can highlight developing asymmetries across sequential exams

The interval cancer rate is a key performance indicator for screening programs and AI systems. It quantifies cancers emerging between scheduled screens per 1,000 negative examinations.

• Acceptable benchmark: 0.5–1.0 interval cancers per 1,000 screens in biennial programs
• Rates above 1.2 per 1,000 trigger program audits
• AI sensitivity is often measured by how many interval cancers it would have retrospectively flagged
• Program sensitivity = screen-detected / (screen-detected + interval cancers) over a defined period

Deep learning models address interval cancers through multiple complementary mechanisms that target both biological and perceptual failure modes.

• Multi-view correlation links subtle findings across CC and MLO views that a radiologist might dismiss in isolation
• Density-invariant feature extraction reduces masking effects in dense tissue
• Temporal subtraction networks highlight interval change between current and prior exams
• Worklist prioritization ensures high-suspicion cases are read first, reducing fatigue-related misses
• Prospective studies show AI-assisted reading can reduce interval cancer rates by 9–15% in screening populations