EUSOBI spotlight: AI turns routine mammograms into near-term breast cancer risk forecasts—reshaping who gets extra screening and when

Artificial intelligence is starting to change what a mammogram can do. It can estimate near-term breast cancer risk, not only detect tumors. At the European Society of Breast Imaging (EUSOBI) annual scientific meeting in Aberdeen, Scotland, researchers presented new evidence. The meeting was held with the British Society of Breast Radiology. For Morocco, this is a realistic path. Mammography is already part of routine care in many sites. The question is how to turn images into better decisions, safely. ### Key takeaways - AI can turn routine mammograms into near-term risk scores that guide extra imaging and earlier follow-up. - Traditional risk models are hard to apply consistently and often miss interval cancers. - Clinical trials and clear guidelines are the main barrier to rollout. - Morocco can pilot imaging-based risk tools, but must invest in governance, validation, and patient communication. ## From detection to near-term risk forecasting Fixed-interval screening is simple, but it is blunt. It can over-screen some women and under-screen others. Mikael Eriksson (Karolinska Institutet) argues for a tighter time horizon. The model targets a clinically actionable window, not lifetime probability. If validated, the same mammogram supports two outputs. One is the radiologist report. The other is a risk score that can trigger a different pathway. ## Why traditional risk models struggle in real clinics Traditional models use age, family history, and other clinical inputs. They can be well-calibrated for broad population risk classes. Eriksson argues they fail in routine screening because inputs are missing or incomplete. He also warns about uneven performance across ethnic subgroups, which can create bias. In his view, the main frictions are practical, not theoretical. Key problems include: - Low uptake in real screening programs. - Heavy reliance on family history instead of general risk. - Difficulty collecting complete inputs during routine care. - Poor identification of women who later develop interval cancers. In Morocco, these frictions can be amplified by fragmented records and time pressure. Imaging-based signals could reduce dependence on perfect questionnaires. ## The AI approach: use the mammogram twice Eriksson's team aims to reuse existing mammography infrastructure for risk assessment. The goal is to predict risk in a near-term window that changes care. In practice, screening services could take three actions. Each one needs a clear protocol and capacity planning: - Offer supplemental imaging to higher-risk women. - Shorten the follow-up interval when appropriate. - Reduce false-positive risk prediction compared with cruder rules. This is not a plug-and-play decision. Workflows must define thresholds, referrals, and accountability. Without that, a score becomes noise. ## Interval cancers: the hard cases AI aims to catch earlier Interval cancers are diagnosed between scheduled screening rounds. Eriksson noted they represent about 15–45% of breast cancers. A narrow-window prediction strategy is designed to reduce that share. It accepts that some extra recalls may occur, and that harm must be managed. For Morocco, interval cancers also intersect with follow-up logistics. Missed appointments and delayed imaging can turn risk into late detection. A risk-guided recall system can help, but only if access barriers are addressed. ## Density, bias, and confounders Breast density increases risk and makes reading harder. Dense tissue can hide tumors on mammography. Eriksson reported that the AI model detects high-risk women regardless of mammographic density. He also argued it outperforms density-only rules for precision screening. The presentation also stressed confounders. AI can learn shortcuts linked to equipment, site practice, or population mix. Several imaging-based risk models are now being validated in multiple screening settings, with promising results. But local calibration and monitoring still matter. ## The missing step before broad rollout: trials, guidelines, and trust Eriksson's main barrier is clinical guidance backed by trials. Strong model performance is not enough on its own. Trials should measure outcomes, not only detection counts. Key endpoints include: - Interval cancer rate and stage at diagnosis. - Recall burden and false positives. - Attendance effects, anxiety, and trust. Guidelines should also cover patient communication. Risk scores change how people perceive their health. In Morocco, communication must work across Arabic, French, and Amazigh contexts. ## What an AI risk pilot could look like in Morocco A Moroccan pilot can start small and stay rigorous. Use a few sites, one workflow, and clear endpoints. Expand only after prospective results. ### Practical uses that fit Moroccan services - Risk-based selection for supplemental ultrasound where clinically justified. - Shorter follow-up intervals for high near-term risk, with pre-defined rules. - Prioritised second reads for exams flagged as higher risk. ### Where Moroccan AI startups can help - Integrate AI into PACS and reporting, with minimal clicks for radiologists. - Build de-identification, consent, and audit tooling for imaging datasets. - Provide monitoring dashboards for drift across devices and hospitals. ### Government actions that unblock scale Morocco already has a data protection framework. Law 09-08 and the CNDP shape how health data can be processed and shared. Beyond privacy, a few policy moves can speed safe adoption: - Create a national evaluation protocol for radiology AI, including bias and calibration checks. - Fund prospective pilots through public hospitals, with transparent reporting of outcomes. - Set procurement rules that require security reviews, audit rights, and clear clinical responsibility. - Support training for clinicians and engineers on AI use, limits, and communication. ## Beyond screening: response prediction is promising, but early Dr. Ritse Mann (Radboud University Medical Center / Netherlands Cancer Institute) focused on AI for response prediction and diagnosis. He called it a golden opportunity, but also a clear work in progress. Image-based AI may improve prediction of pCR (pathologic complete response) compared with clinical features alone. Mann described the effect so far as modest. If the evidence matures, the implications are significant. High-impact possibilities include: - Better selection of less toxic regimens. - Post-treatment imaging AI that helps identify who might avoid surgery or radiotherapy. - Earlier flags for patients likely to need more intensive care. For Morocco, these are medium-term goals. They depend on consistent imaging, outcome tracking, and multidisciplinary oncology pathways. ## A cautious path to impact AI-based near-term risk forecasting could make breast screening more adaptive. It could allocate extra imaging to those most likely to benefit, and reduce interval cancers. Morocco can prepare now with pilots, governance, and local validation. The goal is measurable outcomes and sustained trust, not novelty.