Skip to main content

AI Elevates Tomosynthesis Cancer Detection

August 5, 2025
Image: [image credit]

Mark Hait
Mark Hait, Contributing Editor

Digital breast tomosynthesis entered U.S. screening programs with the promise of sharper lesion visualization and fewer unnecessary recalls. Despite those gains, interval breast cancers, tumors discovered symptomatically after a negative exam and before the next scheduled visit, continue to undermine clinical and financial goals. A retrospective analysis published in Radiology linked an artificial-intelligence algorithm from Lunit to a 32.6 percent detection rate for interval cancers that human readers had missed. That finding, produced by investigators at Massachusetts General Hospital, reframes the strategic conversation around second-reader workflows in breast imaging.

Reframing a Quality Benchmark

Interval cancer rate serves as a surrogate marker for screening quality because few health systems have accrued the decade of mortality data needed to evaluate tomosynthesis directly. The study team examined more than thirteen hundred screening cases and isolated 224 interval cancers. When the algorithm reviewed the original images, it identified seventy-three of those tumors and accurately pinpointed their locations. Notably, algorithm-detected lesions were larger and more likely to involve regional lymph nodes than those still overlooked, suggesting that software support may surface biologically aggressive cancers earlier than current practice allows.

Operational Realities

Radiologists interpreting tomosynthesis scroll through hundreds of high-resolution slices for every patient, a workload that heightens fatigue and inter-reader variability. A 2025 survey by HIMSS reported that two-thirds of imaging leaders now explore artificial-intelligence tools primarily to manage rising image volume. Embedding a validated algorithm as a second reader can mitigate perceptual misses and standardize performance across sites. Effective implementation, however, requires governance policies that define how radiologists weigh software suggestions, document overrides, and track longitudinal quality indicators. Without clear rules, decision support can devolve into alert fatigue or unknown liability exposure.

Financial and Regulatory Stakes

Interval cancers diagnosed at later stages carry higher direct medical costs than early detections. A 2024 analysis in Health Affairs estimated that late-stage breast-cancer treatment triples five-year spending compared with localized disease. Redirecting even a portion of interval cases toward earlier discovery can offset licensing, integration, and maintenance fees associated with artificial-intelligence software. Regulatory expectations are also maturing. The Food and Drug Administration now classifies tomosynthesis support tools as software devices subject to post-market surveillance, while the Centers for Medicare & Medicaid Services has begun issuing coverage language that references AI-assisted diagnostics. Health systems must align deployment with these requirements to avoid reimbursement delays or audit findings.

Equity and Patient Experience

Financial barriers already influence follow-up compliance. Research from Boston Medical Center found that one in five patients would decline recommended diagnostic imaging if personal charges applied, even though initial screening mammography is covered at no cost. An algorithm that reduces unnecessary callbacks without sacrificing sensitivity could limit out-of-pocket exposure for populations prone to screening disparities. The equity dividend is not automatic, however. Model performance must generalize across breast density categories, racial groups, and age cohorts; rigorous site-level monitoring remains essential.

Evidence Gaps That Will Decide Value

Three questions will shape whether artificial-intelligence support becomes a mainstream second reader. First, multicenter prospective trials must confirm that interval-cancer reduction persists when radiologists interact with software in real time rather than retrospectively. Second, comprehensive cost–benefit analyses need to include infrastructure upgrades, change-management labor, and potential malpractice premiums. Third, payer policies will evolve as clinical data mature; early engagement with reimbursement teams can position health systems to capitalize on value-based incentives rather than react to mandates.

Strategic Outlook

Artificial intelligence will not replace skilled breast imagers, yet the technology can augment human detection where perceptual limits persist. A rigorously governed second-reader model that meaningfully lowers interval-cancer incidence would advance both quality metrics and cost stewardship. Health-system leaders who evaluate this tool through a disciplined lens of clinical efficacy, operational feasibility, financial impact, and equity will be positioned to convert promising research into durable gains for patients and payers alike.