AI Shows Promise for Detecting Interval Breast Cancers

Artificial intelligence could help identify some breast cancers that emerge between routine mammography rounds, particularly cancers that were retrospectively visible on earlier images but not detected at the time. A review published in the Journal of Breast Imaging assesses current evidence on AI and interval cancers, which are diagnosed after a negative screening examination and before the next scheduled screen. These cancers are associated with later-stage diagnosis, poorer prognosis and aggressive molecular subtypes. The evidence suggests potential for earlier detection and risk prediction, but current findings remain difficult to compare and have not yet established a measurable reduction in interval cancer rates in routine prospective practice. 

Must Read: AI Challenges in Mammography PGMI Scoring

AI Performs Best When Earlier Signs Are Visible 

Interval cancers are not a single group. Some abnormalities were present on the previous mammogram but were missed or only subtly visible, while others were genuinely occult or had not yet developed when screening took place. This distinction matters because AI has consistently performed better in retrospective analyses when the later cancer had visible features on the earlier examination. Radiological classification itself is inconsistent, with systems distinguishing categories such as false negatives, minimal signs, occult cancers and true interval cancers in different ways. Review methods also vary, limiting comparisons between programmes. 

Across varied screening programmes, AI was more likely to identify retrospectively visible cancers than nonvisible cancers, with performance generally improving as mammographic visibility increased. In a Swedish study, AI correctly localised 58% of retrospectively visible interval cancers, corresponding to a potential 19% reduction at screening. US studies using annual screening and digital breast tomosynthesis also found high exam-level flagging rates, although lesion-specific localisation was markedly lower. 

That gap is important when interpreting reported performance. Exam-level analysis can count a case as detected even when the AI signal does not correspond to the later cancer site, whereas lesion-specific analysis verifies that the algorithm identified the relevant abnormality. AI has also flagged some mammographically occult cancers, but the clinical response to a high AI score without a clearly visible lesion remains uncertain. 

Prediction May Extend Beyond Visible Lesions 

AI is also being studied as a way to estimate the risk of future interval cancer before a lesion becomes clinically apparent. Large European cohorts found higher AI risk scores in the breast that later developed interval cancer, including on examinations that appeared normal at the time. In one analysis, the proportion assigned the highest malignancy score increased as diagnosis approached, rising from 23.1% several examinations earlier to 39.4% on the screening examination immediately before diagnosis. 

Other approaches have combined lesion-detection scores with mammographic texture or breast-density information. These combined models improved risk assessment for future interval cancers compared with the AI component alone. Such findings suggest that algorithms may detect subtle radiographic patterns associated with later cancer development even when a discrete lesion is not visible. Even so, AI generally performs better for screen-detected cancers than for interval cancers, which are less likely to show suspicious mammographic features. 

The practical meaning of those scores is unresolved. Supplemental imaging could provide one response to elevated risk, but risk-adapted screening would have to account for timing, access, cost, capacity, false positives and overdiagnosis. The optimal timing of supplemental imaging relative to mammography also remains uncertain. Prospective evaluation is therefore needed before risk scores without a visible abnormality can be translated into routine management. 

Prospective Evidence Remains Limited 

Most estimates of potential interval cancer reduction come from retrospective studies or simulated reading workflows, often in European programmes using digital mammography, double reading and longer screening intervals. Differences in interval cancer definitions, radiological classification, screening technology, reading protocols and AI algorithms make direct comparisons difficult. Evidence from annual screening, single-reader settings and digital breast tomosynthesis remains comparatively limited. Commercial systems also differ in intended use, supported imaging modalities, training cohorts and output types, so findings from one algorithm or version may not transfer directly to another. 

One large randomised controlled trial has provided prospective evidence. In the Mammography Screening with Artificial Intelligence trial, AI-supported screening produced a non-inferior interval cancer rate compared with standard double reading, while sensitivity was higher and specificity was equivalent. The AI arm also had fewer interval cancers and more favourable tumour characteristics, but the trial was designed to assess non-inferiority rather than superiority, limiting conclusions about the absolute reduction in interval cancer rates. 

Further limitations include retrospective enriched datasets, incomplete lesion-specific verification and potential verification bias when positive AI findings on previously negative examinations were never worked up diagnostically. Long-term outcomes and postmarket surveillance are also limited. Premarket documentation does not include interval cancer rates. Differences between algorithm versions, study designs, reference standards and outcome definitions further restrict estimates of population-level impact. 

AI can identify a meaningful subset of interval breast cancers, especially when earlier mammograms contain visible or subtle findings, and risk models may detect patterns associated with future cancers before clinical presentation. However, evidence remains dominated by retrospective studies conducted in heterogeneous screening environments, while prospective proof of a reduction in interval cancer rates is limited. More rigorous evaluation with longitudinal follow-up is needed across varied screening settings, including assessment of recall rates, radiologist workload, downstream outcomes and management of high-risk AI findings without a visible lesion. 

Source: Journal of Breast Imaging 

Image Credit: iStock