Artificial Intelligence in Breast Imaging: Reducing Misses, Refining Treatment

Artificial intelligence (AI) has been a popular source of discussion in breast imaging, with many questions being raised about its potential integration into the regular workflow of a breast radiologist. Recently developed AI tools claim to enhance diagnostic accuracy and personalize patient care. Potential areas for AI assistance include breast cancer screening, with false negative rates from 0.8 to 2.1 per 1000 examinations being reported in the literature, or in the analysis of large databases to help stratify the treatment of specific subtypes of breast malignancies. From helping spot subtle cancers on mammograms to characterizing tumor subtypes for tailored therapies, three recently published studies offer an up-to-date survey of AI’s potential capabilities in the diagnostic and therapeutic landscape of breast radiology.

AI Flags Missed Cancers in Screening Mammographies

In a Journal of Breast Imaging study (1), researchers evaluated a commercially available AI algorithm on real-world screening mammograms to see if it could catch breast cancers initially missed by radiologists. The AI system identified roughly half of the false-negative cancers on the original screening images (detecting 54% of missed tumors in 2D mammography and 40% in 3D tomosynthesis). Notably, the lesions flagged by AI tended to be invasive luminal A subtype tumors in women with dense breasts (category C or higher), which are often difficult to discern on mammogram. The algorithm flagged these malignancies a median of about 8–9 months earlier than they were detected in clinical practice. This study suggests that incorporating AI as a second reader during screening could decrease interval cancers by alerting radiologists to subtle findings that might otherwise be overlooked. While getting a second read has been shown to improve diagnostic accuracy, it is often not feasible given the high imaging volumes and demanding workflows radiologists face. AI, functioning as a consistent second reader, may offer a practical way to enhance diagnostic accuracy for breast imaging patients without significantly adding to the radiologist’s workload.

AI Enhances Precision for Triple-Negative Breast Cancer

Another study published in Clinical Breast Cancer (2) reviewed how AI can advance the diagnosis and management of triple-negative breast cancer (TNBC) especially in the era of precision medicine, a subset of treatment that is individualized to a patient’s specific genetic profile, lifestyle, and environment. Recent AI-driven models are now able to stratify TNBC into biologically and prognostically distinct subclasses, which is a critical step toward more personalized care. By analyzing imaging features and integrating molecular biomarkers, these tools can help identify which patients are most likely to benefit from specific therapies. Some algorithms can now predict pathologic complete response to neoadjuvant chemotherapy or assess likelihood of response to immunotherapies like checkpoint inhibitors. This predictive capability may ultimately allow clinicians to tailor treatment plans based not only on static receptor status but also on dynamic tumor behavior. In this way, AI is helping shift TNBC management from a one-size-fits-all approach to a more nuanced, response-guided strategy. This shift could be especially impactful, as TNBC is known to be highly chemo-resistant and carries the lowest 5-year survival rate among all breast cancer subtypes. Additionally, tailoring chemotherapy to TNBC subtypes can reduce the toxicity experienced by healthy cells from the standard treatment regimen, often a potent combination of anthracyclines and taxanes. By identifying which subclass a tumor falls into and predicting its likely response to specific therapies, AI can help reduce unnecessary exposure to these agents and direct treatment teams towards a more efficient therapeutic plan.

Standalone AI Excels in Post-Mastectomy Surveillance

A recent study in Radiology (3) examined whether AI could independently read second breast cancer surveillance mammograms in women previously treated with unilateral mastectomy, with the purpose of monitoring the remaining breast for cancer occurrence. In a cohort of 4,184 women with a prior breast cancer, a standalone AI system detected more contralateral cancers than radiologists did, yielding a higher cancer detection rate (17.4 vs 14.6 per 1000 exams) and sensitivity (65.8% vs 55.0%). Notably, the AI algorithm caught 16 out of 50 cancers (32%) that the radiologists missed on these follow-up mammograms. However, the AI had a slightly lower specificity (91.5% vs 98.1%), generating more false-positive alerts than radiologists. While some tumors were still missed by both the AI and radiologists, the findings demonstrate AI’s potential as a valuable second reader in high-risk surveillance imaging. It can pick up a significant fraction of cancers that might otherwise escape notice, albeit with an increase in unnecessary recalls. The authors suggest using AI in an interactive way alongside radiologists to capitalize on AI’s sensitivity while minimizing false positives.

The Future of Timely Detection and Personalized Care

Together, these studies highlight how AI, as a second reader, could be a powerful tool for improving diagnostic accuracy without increasing workload. In a field where additional reads are often ideal but logistically unfeasible due to time and volume constraints, AI presents a scalable solution to enhance detection quality across large patient populations. It can serve as a practical second set of eyes, flagging early-stage and subtle cancers that may be inadvertently missed, resulting in earlier diagnoses for at-risk patients. Additionally, AI’s ability to sift vast pathology and imaging data is helpful in advancing personalized care in the age of precision medicine, from pinpointing TNBC characteristics on MRI to predicting which patients will respond to specific therapies. With the possibility of integrating AI into clinical workflows, breast imaging may move towards a future where fewer cancers are missed, and each patient’s treatment can be more finely tailored to the biology of their disease.

However, despite its promise, AI is not without limitations. Most models still require external validation across diverse populations and practice settings. There is also a need for clearer regulatory pathways and integration protocols to standardize AI deployment in clinical environments. Furthermore, many systems remain “black boxes,” offering predictions without clear insight into their decision-making processes, posing challenges for clinician trust and accountability. If radiologists are unable to easily assess and verify the rationale behind an AI system’s output, it may lead to uncertainty and delays in clinical workflow. Continued research, including prospective studies and multi-institutional trials, will be essential to evaluate long-term outcomes, ensure equity in performance across subgroups, and identify best practices for integration.

As the field continues to evaluate these emerging technologies, the importance of collaboration between clinical expertise and AI tools becomes increasingly evident. This synergy of radiologist expertise and AI assistance holds promise for improving both the early detection of breast cancer and the individualization of therapy for patients.