Eye on AI
By Rebecca Montz, EdD, MBA, CNMT, PET, RT(N)(CT), NMTCB RS
Radiology Today
Vol. 26 No. 7 P. 18

AI is ushering in a new era of precision and efficiency in breast imaging.

AI is no longer a futuristic concept in breast imaging. It is a present-day force, actively transforming how breast cancer is detected, diagnosed, and managed. As the demand for more efficient, accurate, and equitable screening continues to grow, AI is emerging not as a replacement for radiologists but as a powerful tool that enhances their capabilities.

From digital breast tomosynthesis (DBT) to MRI, AI-driven technologies are streamlining workflows and improving diagnostic precision. These tools are designed to recognize complex patterns, highlight areas of concern, reduce false positives, and assist in triaging exams. By doing so, they help radiologists make faster and more confident decisions while managing increasingly complex imaging data.

What once seemed like science fiction— machines supporting human interpretation with predictive precision—is now becoming standard practice in clinics and imaging centers across the country. At the same time, the rise of AI brings important considerations around safety, bias, transparency, and the need for thoughtful integration into clinical environments. From novel anomaly detection models used in breast MRI to FDA-cleared AI tools improving DBT interpretation, there are opportunities and challenges that come with adopting AI in the fight against breast cancer.

Raising the Bar for DBT
AI is emerging as a powerful tool for DBT, particularly in efforts to improve early cancer detection and reduce interval cancer rates. RSNA recently highlighted promising findings from the study, “AI to Reduce the Interval Cancer Rate of Screening Digital Breast Tomosynthesis,” led by Manisha Bahl, MD, an associate professor of radiology at Harvard Medical School and Breast Imaging Division quality director at Massachusetts General Hospital. The research evaluated the FDA-cleared Lunit INSIGHT DBT tool and its ability to retrospectively detect interval cancers, which are diagnosed between routine screenings.

The study analyzed 224 interval cancer cases and found that the AI tool correctly localized nearly one-third of the cancers. These findings represent a meaningful advancement, given that interval cancers often represent aggressive disease types and are associated with poorer outcomes. The AI model also performed strongly in screening-detected cases, accurately localizing more than 84% of true-positive cancers, and it correctly categorized a large proportion of true-negative and falsepositive exams as negative. These findings suggest that such tools could help reduce the interval cancer rate, a key surrogate for long-term screening effectiveness.

Lunit INSIGHT DBT is designed to function as a concurrent or second reader, highlighting suspicious regions and assigning malignancy likelihood scores that radiologists can incorporate into their workflow, either during or after their independent interpretation. When properly implemented, this approach supports radiologists in identifying subtle findings that might otherwise go unnoticed, enhancing sensitivity without significantly disrupting workflow.

According to Bahl, successful integration of AI tools such as this requires thoughtful planning. Technical integration with PACS or mammography reading workstations is essential, as are training and user adoption strategies that foster trust and consistent usage. While the AI tool can act as a second pair of eyes in high-volume clinical settings, its use must be balanced to avoid increased false positives or undue influence on radiologists’ decision making.

Overall, the research underscores both the potential and the complexity of incorporating AI into routine breast cancer screening. With continued validation and careful integration, AI-enhanced DBT may help clinicians detect cancers earlier, optimize workflows, and ultimately improve patient outcomes.

Influencing Radiologist Behavior
At Radboud University Medical Center in the Netherlands, researchers have uncovered how AI tools influence radiologist behavior during the interpretation of mammography images. Jessie JJ Gommers, MSc, the first author of the study, “Influence of AI Decision Support on Radiologists’ Performance and Visual Search in Screening Mammography,” led the research team in evaluating how AI-driven decision support affects visual attention and diagnostic performance.

Using an eye-tracking system, the study monitored how 12 radiologists examined mammograms from 150 women, half with confirmed breast cancer, with and without AI assistance. The presence of AI not only improved overall cancer detection accuracy but also altered how radiologists distributed their attention. With AI support, radiologists spent significantly more time reviewing regions that were flagged by the algorithm and less time on unmarked areas. These flagged areas often corresponded to actual cancerous lesions, underscoring AI’s ability to guide attention toward clinically meaningful findings.

The study further revealed that radiologists were more likely to adjust their diagnostic decisions based on the level of suspicion indicated by the AI. High malignancy scores often prompted more cautious review, especially in complex or subtle cases, while low scores helped radiologists move more quickly through clearly benign or normal findings. This dynamic interaction shows that AI is not just a passive aid but an active influence on how radiologists prioritize and process imaging information.

However, this behavioral shift also introduces risk. If a lesion is not highlighted by AI, radiologists may inadvertently overlook it, especially in high-volume or time-constrained environments. The findings emphasize the importance of ensuring that radiologists remain critical interpreters and do not become overly dependent on algorithmic suggestions. Gommers and her team note that how and when AI input is presented—whether before, during, or after the radiologist’s initial impression—can influence diagnostic outcomes and warrants further research.

By leveraging eye-tracking data, the study provides more than just performance metrics, it offers a window into radiologists’ decision making processes. This data could inform future AI interface design, ensuring that tools support radiologists without narrowing their focus too much. It also highlights the need to educate radiologists on maintaining independent judgment while interpreting AI outputs.

Ultimately, the research led by Gommers underscores the dual potential of AI to enhance both the accuracy and efficiency of breast cancer screening. But it also serves as a caution: Thoughtful integration, clinician accountability, and ongoing evaluation remain essential to maximizing AI’s benefits without compromising diagnostic vigilance.

Anomaly Detection
Breast MRI is a powerful imaging modality, known for its high sensitivity in detecting breast cancer, particularly in women with dense breast tissue or elevated risk. However, the complexity and volume of MRI data, coupled with longer scan and interpretation times, have created a strong need for tools that can improve efficiency without compromising diagnostic performance. Recent research led by Savannah C. Partridge, PhD, FSBI, FAIMBE, a professor of radiology at the University of Washington and Fred Hutchinson Cancer Center, explores the potential of an AI model designed specifically for breast MRI using an anomaly detection approach.

Unlike conventional classification models that aim to distinguish between cancerous and noncancerous cases based on balanced datasets, anomaly detection models learn to recognize the characteristics of normal or benign cases. By understanding what “normal” looks like, the model is better equipped to flag subtle abnormalities that may indicate malignancy, even in low prevalence screening populations where cancer is rare. This approach aligns more closely with real-world screening scenarios, where the majority of breast MRI exams are negative or benign.

Partridge coauthored the study, “Cancer Detection in Breast MRI Screening via Explainable AI Anomaly Detection,” which tested the model on both internal and external datasets. The AI-generated heat maps effectively highlighted areas of concern, with the flagged regions closely matching biopsy-proven malignancies identified by radiologists. While the model was not directly compared with radiologist performance in this study, early results suggested the tool may offer higher specificity with somewhat lower sensitivity when used independently. However, the model’s greatest potential lies in augmenting radiologists’ performance, acting as a triage tool that helps direct attention to suspicious regions while reducing time spent on clearly benign findings.

A significant advantage of this anomaly detection model is its compatibility with abbreviated MRI protocols. By requiring only minimal image data, it offers the possibility of reducing scan and interpretation times, making breast MRI more accessible for patients and less burdensome for imaging departments. In busy clinical settings, such tools could streamline workflows by quickly excluding normal scans from further review and escalating suspicious cases for priority interpretation or additional workup.

For successful clinical integration, researchers emphasize the importance of determining how best to incorporate these AI tools into radiologist workflows. Questions remain about when the AI output should be presented, whether at the beginning of the interpretation process or after an initial human review, and how it affects radiologist decision making in real-world settings. Ongoing research will help clarify the optimal methods for presentation and interaction to ensure AI enhances, rather than distracts from, clinical performance.

The study by Partridge and colleagues marks a pivotal shift in AI development for breast imaging, moving beyond basic classification toward more sophisticated, context-aware tools. Anomaly detection models may help close the gap between early detection and efficient workflow, especially in resource-constrained environments where access to advanced imaging and expert interpretation can be limited. With continued validation in larger, diverse populations and prospective clinical trials, these tools could reshape the role of MRI in breast cancer screening and diagnosis.

From Detection to Workflow
From the industry side of breast imaging, Hologic offers AI technologies designed to enhance both diagnostic accuracy and workflow efficiency. Mark Horvath, president of breast and skeletal health solutions at Hologic, says the company’s evolving AI strategy includes tools such as the Genius AI Detection suite, which includes the Genius AI Detection PRO solution.

In FDA-reviewed clinical validation studies, radiologists using the Genius AI Detection 2.0 platform showed measurable improvements in sensitivity and specificity. The latest PRO version incorporates prior mammograms into the analysis, enabling longitudinal comparisons that enhance specificity and positive predictive value, while reducing unnecessary recalls.

Horvath emphasizes AI’s impact on radiologist workload. Tools like the Genius AI Detection PRO solution offer functionalities such as patient triaging and automated report population, which help reduce perceived fatigue and support more efficient workflows. In Hologic’s own clinical research, radiologists reported significantly lower fatigue when using the PRO platform.

Despite these advancements, Horvath acknowledges key challenges, especially around workflow integration and data diversity. The performance of AI tools is highly dependent on the quality and representativeness of training datasets. To address this, Hologic conducted large-scale evaluations of its Genius AI Detection 2.0 platform on more than 7,500 DBT cases across diverse racial and ethnic groups, demonstrating consistent performance among Asian, Black, Hispanic, and white patients.

Trust and explainability remain critical to AI adoption. Hologic’s AI tools are designed with visual outputs, including color-coded lesion scores and confidence indicators. The aim is to keep radiologists in full control of the diagnostic process, with AI serving as a support system rather than a replacement.

Clinical readiness, according to Horvath, depends not only on regulatory clearance but also on integration with existing IT infrastructures. The Genius AI Detection PRO solution addresses interoperability by offering an all-in-one application compatible with Hologic’s SecurView, as well as third-party workstations.

AI as a Partner in Care
The case for AI in breast imaging is no longer hypothetical. With demonstrable improvements in detection, workflow efficiency, and radiologist satisfaction, AI tools represent a new standard in breast imaging innovation.

“The transformation of breast imaging isn’t coming—it’s here,” Horvath says.

As AI continues to evolve, one message remains consistent across research, clinical practice, and industry: AI must serve as a partner, not a replacement. Bahl, Gommers, and Partridge agree on the importance of maintaining clinical oversight, warning against overreliance on algorithmic outputs. Horvath echoes this consensus, reinforcing that AI is most powerful when it augments, rather than replaces, radiologist expertise.

The path forward depends on thoughtful collaboration between data scientists, engineers, clinicians, and regulators. This multidisciplinary effort is essential not only for refining algorithm performance but also for ensuring transparency, equity, and clinical relevance in diverse real-world settings.

As AI tools continue to evolve, their true value lies in amplifying human judgment, enhancing diagnostic confidence, improving workflow efficiency, and broadening access to timely, high-quality breast imaging. Rather than disrupting care, AI is helping reshape it for the better.

— Rebecca Montz, EdD, MBA, CNMT, PET, RT(N) (CT), NMTCB RS, has worked at the Mayo Clinic Jacksonville and University of Texas MD Anderson Cancer Center in Houston as a nuclear medicine and PET technologist.