By Billy Candra 
A groundbreaking new study led by investigators at the UCLA Health Jonsson Comprehensive Cancer Center suggests that artificial intelligence (AI) could revolutionize breast cancer screening by helping to detect interval breast cancers – those that emerge between routine screenings – before they become more advanced and challenging to treat. This could potentially lead to significantly better screening practices, earlier therapeutic interventions, and ultimately, improved patient outcomes, marking a pivotal advancement in the ongoing battle against breast cancer.
Understanding the Challenge of Interval Breast Cancers
Interval breast cancers represent a critical challenge in oncology. Unlike screen-detected cancers, which are identified during scheduled mammograms, interval cancers manifest symptomatically (e.g., a palpable lump) in the period between a negative screening mammogram and the next scheduled screening. These cancers are often more aggressive, grow rapidly, and are associated with a poorer prognosis compared to screen-detected cancers due to their later presentation. They account for a significant proportion, often ranging from 20% to 30%, of all breast cancer diagnoses, despite regular screening efforts. The difficulty in detecting them stems from several factors, including their rapid growth rate, the possibility of being genuinely "occult" or invisible on mammograms at the time of screening, or presenting with very subtle signs that are easily overlooked by the human eye. The emotional and physical toll on patients diagnosed with interval cancers, who believed they were cancer-free after a recent screening, can also be particularly severe.
The UCLA Study: AI as a "Second Set of Eyes"
Published in the esteemed Journal of the National Cancer Institute, the UCLA study provides compelling evidence that AI possesses the capability to identify "mammographically-visible" types of interval cancers at an earlier stage. The AI accomplished this by flagging suspicious areas at the time of the initial screening mammogram, even when human radiologists did not detect them. These include tumors that, while present on mammograms, were not initially recognized by radiologists, or those exhibiting extremely subtle radiological signs that were either too faint to register or arguably below the threshold of human perception.
Researchers involved in the study estimate that integrating AI into current screening protocols could lead to a substantial reduction—up to 30%—in the number of interval breast cancers. Dr. Tiffany Yu, an assistant professor of Radiology at the David Geffen School of Medicine at UCLA and the study’s first author, underscored the profound implications of this finding. "This finding is important because these interval cancer types could be caught earlier when the cancer is easier to treat," Dr. Yu stated. "For patients, catching cancer early can make all the difference. It can lead to less aggressive treatment, such as avoiding extensive surgery or chemotherapy, and significantly improve the chances of a better outcome and long-term survival."
Methodology and Key Findings
The retrospective study meticulously analyzed an extensive dataset comprising nearly 185,000 past mammograms collected between 2010 and 2019. This dataset included both digital mammography (DM), commonly known as 2D mammography, and digital breast tomosynthesis (DBT), often referred to as 3D mammography. From this vast pool of data, the research team specifically focused on 148 cases where a woman had been subsequently diagnosed with interval breast cancer.
A critical step in the study involved radiologists meticulously reviewing these 148 cases to ascertain the reasons why the cancer was not initially detected. To systematically categorize these missed diagnoses, the research team adapted a classification system predominantly used in European studies. This system breaks down interval cancers into several categories:
- Missed reading error: The cancer was visible but overlooked by the radiologist.
- Minimal signs-actionable: Subtle signs were present that, in retrospect, should have prompted further investigation.
- Minimal signs-non-actionable: Very faint signs were present, but their significance was difficult to interpret at the time.
- True interval cancer: The cancer was genuinely not visible on the prior mammogram but developed rapidly thereafter.
- Occult: The cancer was truly invisible on the mammogram, even upon retrospective review.
- Missed due to a technical error: Issues with image acquisition or processing contributed to the miss.
Following this detailed human review, the researchers applied a commercially available AI software, named Transpara, to the initial screening mammograms that were performed before the interval cancer diagnosis. The objective was to determine if the AI could detect the subtle signs of cancer that had been missed by human radiologists during their initial screenings, or at the very least, flag them as suspicious. The AI tool assigned a cancer risk score to each mammogram on a scale of 1 to 10. A score of 8 or higher was designated as a "flagged" or potentially concerning result, indicating a higher probability of malignancy. The AI’s ability to re-evaluate these previously missed cases provided a robust test of its diagnostic capabilities.
Bridging the Atlantic: U.S. vs. European Screening Practices
While similar research exploring the utility of AI in detecting interval breast cancers has been conducted in Europe, the UCLA study stands out as one of the first to specifically investigate this application within the unique context of the United States healthcare system. This distinction is crucial due to fundamental differences in screening practices between the two continents.
In the U.S., the majority of mammograms are performed using digital breast tomosynthesis (DBT), or 3D mammography, which provides a more detailed, layered view of breast tissue, reducing the impact of overlapping structures compared to traditional 2D mammography. Furthermore, patients in the U.S. are typically recommended for annual screenings, emphasizing early and frequent surveillance.
Conversely, European screening programs have historically relied more heavily on digital mammography (DM), or 2D mammography. Moreover, screening intervals in Europe are generally longer, with patients typically undergoing mammograms every two to three years. These differences in imaging technology and screening frequency could significantly influence the performance and utility of AI systems. A system trained primarily on 2D mammograms from biennial screenings might perform differently when applied to 3D mammograms from annual screenings. The UCLA study’s focus on U.S. data, particularly including DBT, therefore offers invaluable insights directly applicable to the American healthcare landscape and contributes to a global understanding of AI’s potential in diverse clinical settings.
The Nuances of AI Performance: Promise and Pitfalls
While the study yielded undeniably exciting results, it also brought to light important considerations regarding AI’s accuracy and limitations that necessitate further exploration in real-world clinical environments. Dr. Hannah Milch, an assistant professor of Radiology at the David Geffen School of Medicine and the senior author of the study, candidly addressed these nuances. "While we had some exciting results, we also uncovered a lot of AI inaccuracy and issues that need to be further explored in real-world settings," Dr. Milch noted.
One particularly insightful finding concerned occult cancers—those truly invisible on mammography, even upon retrospective review. Despite their inherent invisibility to the human eye, the Transpara AI tool remarkably flagged 69% of the initial screening mammograms that subsequently developed into occult cancers. This suggests AI’s capacity to detect extremely subtle, perhaps pre-mammographic, indicators of risk that are entirely imperceptible to humans. However, the study also revealed a significant challenge: when researchers zoomed in to assess the specific areas on the images that the AI marked as suspicious, the AI demonstrated a lower localization accuracy, pinpointing the actual cancer only 22% of the time. This disparity highlights a crucial distinction: AI can be effective at identifying risk or suspicious regions, but its precision in localizing the exact pathology still requires refinement.
These findings underscore a fundamental principle: AI is a powerful diagnostic aid, not a standalone replacement for human expertise. Its ability to flag potential concerns, even with imperfect localization, can serve as a valuable prompt for radiologists, encouraging a second, more intensive review of flagged images. However, the high rate of flagging occult cancers without precise localization also raises questions about potential increases in false positives and the subsequent need for additional, potentially invasive, follow-up procedures like biopsies, which carry their own risks and costs. Larger prospective studies are therefore essential to understand how radiologists would practically integrate AI into their workflow, and to address key operational questions. How, for instance, should clinicians manage cases where AI flags suspicious areas that are not discernible to the human eye, especially when the AI itself isn’t always accurate in pinpointing the precise location of the abnormality? Balancing the benefits of early detection with the risks and anxieties associated with false positives will be a critical aspect of AI implementation.
The Evolving Landscape of AI in Medical Imaging
The integration of AI into medical imaging represents a significant evolutionary step from earlier computer-aided detection (CAD) systems. CAD systems, introduced decades ago, were primarily rule-based algorithms designed to highlight calcifications or masses, often resulting in high false-positive rates that could lead to alert fatigue for radiologists. Modern AI, particularly deep learning models, operates differently. These systems are trained on vast datasets of medical images, learning complex patterns and features associated with diseases, often surpassing human capabilities in pattern recognition for specific tasks.
The journey of AI in radiology has been marked by rapid technological advancements. From early promise to current sophisticated algorithms, AI is transforming how medical images are interpreted. However, its adoption is not without hurdles. Regulatory bodies, such as the FDA in the U.S., must rigorously evaluate AI tools for safety and efficacy. Integrating AI into existing clinical workflows requires careful planning, ensuring seamless interaction with Picture Archiving and Communication Systems (PACS) and Electronic Health Records (EHRs). Furthermore, concerns regarding data bias—where AI models trained on unrepresentative datasets might perform poorly on diverse patient populations—and the ‘explainability’ of AI decisions remain central to fostering trust and widespread adoption.
Implications for Clinical Practice and Patient Care
The findings from the UCLA study have profound implications across several facets of clinical practice and patient care:
- Enhanced Radiologist Workflow: AI, acting as an intelligent "second reader," could significantly enhance the efficiency and accuracy of radiologists. By flagging subtle or easily missed anomalies, AI can help reduce diagnostic fatigue, especially in high-volume screening environments. This partnership could lead to more consistent interpretations and fewer missed cancers, allowing radiologists to focus their expertise on complex cases and critical decision-making.
- Personalized Screening Strategies: The ability of AI to assess risk at the individual mammogram level opens doors for more personalized screening approaches. Patients deemed by AI to be at higher risk, even with an ostensibly normal human reading, might be recommended for earlier follow-up or additional imaging modalities, such as MRI. Conversely, those consistently identified as very low risk might benefit from adjusted screening intervals, optimizing resource allocation.
- Patient Empowerment and Outcomes: For patients, the promise of earlier detection translates directly into potentially less aggressive treatment regimens. Catching cancer when it is smaller and localized often means less invasive surgery, fewer rounds of chemotherapy or radiation, and a significantly higher chance of survival and a better quality of life post-treatment. This proactive approach could also alleviate the immense anxiety associated with waiting for screening results and the fear of an interval cancer diagnosis.
- Healthcare Economics: While the initial investment in AI infrastructure and software might be substantial, the long-term economic benefits could be considerable. Earlier detection often leads to simpler, less costly treatments compared to managing advanced-stage cancers. Reduced morbidity and mortality rates also contribute to a healthier, more productive population, indirectly benefiting the broader economy. However, the potential for increased false positives, leading to additional diagnostic workups, needs careful cost-benefit analysis.
Future Directions and Remaining Questions
Despite the encouraging results, the UCLA researchers, like many in the field, emphasize that this study represents an important step, not the final destination. Critical questions remain, necessitating further rigorous investigation.
- Larger Prospective Studies: The immediate need is for larger, prospective studies conducted in diverse real-world settings. These studies would observe how radiologists actually integrate AI into their daily practice, evaluate its impact on screening outcomes, and precisely quantify reductions in interval cancer rates and false-positive rates. Such studies are crucial for validating the findings on a broader scale and establishing clear clinical guidelines for AI use.
- Optimizing AI Algorithms: Ongoing research is vital to refine AI algorithms, aiming to improve localization accuracy—so that when AI flags an area, it can precisely pinpoint the abnormality—and to reduce the incidence of false positives. This involves developing more sophisticated models, training them on even larger and more diverse datasets, and exploring multi-modal AI that combines mammographic data with other patient information (e.g., genetic markers, clinical history).
- Radiologist Training and Integration: Effective integration of AI will require training radiologists not just on how to operate the software, but on how to interpret its output critically, understand its limitations, and leverage its strengths. This includes understanding when to trust AI’s flags, when to override them, and how to communicate complex AI-driven findings to patients.
- Ethical Considerations: The ethical dimensions of AI in healthcare are paramount. Questions of accountability (who is responsible when AI makes a mistake?), data privacy and security, and the potential for perpetuating or even amplifying existing biases in healthcare (if training data is not diverse) must be thoroughly addressed. The balance between AI-driven efficiency and the human element of compassionate care is also a key consideration.
Dr. Yu reiterated the overarching vision for AI’s role: "While AI isn’t perfect and shouldn’t be used on its own, these findings support the idea that AI could help shift interval breast cancers toward mostly true interval cancers." This implies a future where fewer cancers are missed due to human oversight or subtle signs, leaving only those that genuinely develop too rapidly or are inherently invisible at the time of screening. "It shows potential to serve as a valuable second set of eyes, especially for the types of cancers that are the hardest to catch early. This is about giving radiologists better tools and giving patients the best chance at catching cancer early, which could lead to more lives saved."
The collaborative effort behind this significant study involved numerous other authors from UCLA, including Dr. Anne Hoyt, Dr. Melissa Joines, Dr. Cheryce Fischer, Dr. Nazanin Yaghmai, Dr. James Chalfant, Dr. Lucy Chow, Dr. Shabnam Mortazavi, Christopher Sears, Dr. James Sayre, Dr. Joann Elmore, and Dr. William Hsu. The vital work was supported in part by generous funding from the National Institutes of Health, the National Cancer Institute, the Agency for Healthcare Research and Quality, and Early Diagnostics Inc., highlighting the broad scientific and institutional backing for this critical area of research. As AI continues to evolve, its promise in transforming breast cancer screening from a reactive process to a more proactive and precise one brings renewed hope for countless patients worldwide.
