By Billy Candra 
A groundbreaking study led by investigators at the UCLA Health Jonsson Comprehensive Cancer Center offers compelling evidence that artificial intelligence (AI) could significantly enhance the early detection of interval breast cancers—those aggressive malignancies that emerge between routine screening appointments. This innovative application of AI holds the promise of transforming current screening protocols, facilitating earlier intervention, and ultimately leading to improved patient prognoses and survival rates. The research specifically targets cancers that, despite being visible on mammograms, are often missed by human radiologists due to their subtle presentation or faint signs, thereby addressing a critical gap in current diagnostic practices.
Understanding the Challenge of Interval Breast Cancers
Breast cancer remains one of the most common cancers among women globally, with early detection being a cornerstone of successful treatment. Routine mammography screenings have been instrumental in reducing breast cancer mortality by identifying tumors at an early, more treatable stage. However, a significant challenge persists in the form of "interval breast cancers." These are cancers diagnosed within the interval between a normal screening mammogram and the subsequent scheduled screening, typically within 12 to 24 months. Unlike screen-detected cancers, interval cancers often present with more aggressive biological characteristics, grow faster, and are associated with a poorer prognosis because they are typically diagnosed at a more advanced stage. They account for an estimated 20-30% of all breast cancers and are a leading cause of patient anxiety and diagnostic uncertainty.
The difficulty in detecting interval cancers stems from several factors. Some are "true interval cancers," meaning they develop rapidly after a negative screening. Others are "missed cancers," where signs were present on the earlier mammogram but were not identified by the interpreting radiologist. These missed cases can be further categorized: some have "minimal signs" that are actionable but easily overlooked, others have "minimal signs" that are arguably below the threshold of human detection, or they might be missed due to technical errors in image acquisition or interpretation. A small proportion are "occult," meaning they are truly invisible on mammograms. The clinical implications of missing such cancers are profound, potentially delaying life-saving treatment and leading to more aggressive therapeutic regimens.
The UCLA Health Study: Methodology and Context
Published in the prestigious Journal of the National Cancer Institute, the UCLA study represents a significant step forward, particularly for the United States, in leveraging AI for breast cancer screening. While similar research has been conducted in Europe, the UCLA team’s work is notable for its focus on U.S. screening practices, which differ considerably from their European counterparts. In the U.S., digital breast tomosynthesis (DBT), often referred to as 3D mammography, is widely used, and annual screenings are common. Conversely, European programs frequently employ 2D digital mammography (DM) and typically screen patients every two to three years. These variations in technology and frequency necessitate tailored research to ensure AI solutions are effective and appropriate for the specific healthcare landscape.
The retrospective study meticulously analyzed a vast dataset comprising nearly 185,000 past mammograms performed between 2010 and 2019, encompassing both DM and DBT images. From this extensive pool, researchers identified 148 cases where a woman was subsequently diagnosed with interval breast cancer. A team of expert radiologists then meticulously reviewed these 148 cases to ascertain why the cancers were not detected during the initial screening. To standardize this review, the study adapted a robust European classification system for interval cancers, categorizing them into: "Missed reading error," "minimal signs-actionable," "minimal signs-non-actionable," "true interval cancer," "occult" (invisible on mammogram), and "missed due to a technical error." This detailed categorization was crucial for understanding the specific types of interval cancers that AI might be best equipped to identify.
Following the human expert review, a commercially available AI software, Transpara, was applied to the initial screening mammograms of these 148 cases. The AI’s task was to determine if it could identify subtle cancerous signs that had eluded radiologists during their initial interpretation or, at the very least, flag these areas as suspicious. The AI tool assigned a cancer risk score ranging from 1 to 10 for each mammogram, with a score of 8 or higher triggering a "potentially concerning" flag. This threshold was critical for evaluating the AI’s sensitivity and specificity in a real-world context.
Key Findings: A Glimpse into AI’s Potential and Current Limitations
The study yielded promising results, suggesting that the integration of AI into current screening practices could potentially reduce the incidence of interval breast cancers by as much as 30%. This projected reduction is particularly significant for "mammographically-visible" types of interval cancers—those that, in hindsight, were discernible on the initial mammogram but were not detected by human radiologists. Such cancers include tumors with very subtle signs that are easily missed due to their faintness or being below the perceptual threshold of the human eye.
Dr. Tiffany Yu, assistant professor of Radiology at the David Geffen School of Medicine at UCLA and the study’s first author, emphasized 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 and improve the chances of a better outcome." This sentiment underscores the core mission of improving patient care and reducing the physical and emotional burden of cancer treatment.
However, the study also provided a realistic assessment of AI’s current capabilities and highlighted areas requiring further refinement. Dr. Hannah Milch, assistant professor of Radiology at the David Geffen School of Medicine and senior author of the study, candidly discussed 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. A prime example involved occult cancers, which are inherently invisible on mammography. Despite their non-visibility, the AI tool flagged 69% of the screening mammograms associated with occult cancers as suspicious. Yet, when researchers examined the specific areas on the images that the AI marked, the tool accurately pinpointed the actual cancer location in only 22% of these cases. This discrepancy indicates that while AI can identify general patterns of concern, its precision in localizing truly occult lesions still needs improvement.
These findings suggest that while AI can serve as a valuable "second set of eyes," it is not yet a standalone diagnostic tool. Its ability to flag general areas of suspicion, even for occult cancers, could be beneficial, but the lack of precise localization for such cases presents a practical challenge for radiologists. How would a radiologist proceed if an AI flags a region as suspicious but offers no clear visual correlate? This question points to the necessity of further research into how radiologists would integrate and trust AI’s suggestions, particularly when faced with AI detections that lack immediate human verification.
The Evolution of AI in Medical Imaging and Broader Implications
The application of AI in medical imaging is not a new concept, but its sophistication has rapidly advanced. Early AI systems, often rule-based, were limited in their ability to handle the complexities of biological data. The advent of deep learning and neural networks, however, has revolutionized the field. These advanced AI models can learn intricate patterns from vast datasets, enabling them to assist in tasks ranging from image segmentation and disease classification to predicting patient outcomes. In breast imaging, AI is being explored for various applications, including risk assessment, density classification, and even synthetic mammography. The UCLA study builds upon this foundation, specifically targeting the nuanced problem of interval cancer detection.
The broader implications of this research are multi-faceted. From a clinical perspective, incorporating AI could lead to a paradigm shift in screening protocols. Radiologists might use AI as a pre-screening tool to prioritize cases, or as a concurrent reader to flag suspicious areas they might have missed. This could potentially reduce radiologist burnout, improve diagnostic accuracy, and standardize interpretation across different practitioners. For patients, the prospect of earlier detection, particularly for aggressive interval cancers, means a greater likelihood of less invasive treatments, such as lumpectomy instead of mastectomy, and a higher chance of long-term survival.
Economically, the impact could also be substantial. Earlier detection often translates to less costly treatments. Preventing cancers from progressing to advanced stages could reduce the overall burden on healthcare systems, both in terms of direct treatment costs and the indirect costs associated with lost productivity and long-term care. However, the initial investment in AI software, infrastructure, and radiologist training must also be considered.
Ethically, the deployment of AI in such critical diagnostic roles raises important questions. Issues of algorithmic bias, data privacy, and accountability for AI-related errors need careful consideration. The study’s findings on AI inaccuracies, particularly for occult cancers, underscore the importance of maintaining human oversight and developing robust regulatory frameworks. The Food and Drug Administration (FDA) in the U.S. has already begun to establish pathways for the approval of AI-powered medical devices, but ongoing validation and real-world performance monitoring will be crucial.
Challenges and Future Directions
Despite its promise, the path to widespread AI integration in breast cancer screening is not without hurdles. The UCLA researchers themselves highlighted the need for larger prospective studies. These studies would involve applying AI in real-time clinical settings to understand how radiologists actually interact with the technology, how it influences their decision-making, and what impact it has on patient outcomes in a live environment. Key questions remain: How will radiologists manage AI flags for areas that are not visibly cancerous to the human eye? How will patient anxiety be managed if AI identifies subtle, potentially non-actionable findings? What is the optimal workflow for integrating AI without overwhelming radiologists with false positives or creating diagnostic fatigue?
Furthermore, the continuous evolution of AI algorithms means that ongoing research and development will be essential. Future AI models may achieve higher levels of precision in localizing occult lesions or in distinguishing benign findings from malignant ones. Training data sets will need to be diverse and representative to prevent algorithmic biases that could disproportionately affect certain patient populations.
A Collaborative Future for AI and Human Expertise
Dr. Yu’s concluding remarks encapsulate the balanced perspective offered by the study: "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 vision suggests a future where the majority of missed cancers are minimized, leaving primarily those truly aggressive tumors that emerge rapidly between screenings. "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 study emphatically supports a collaborative model where AI acts as an intelligent assistant, augmenting human expertise rather than replacing it. By providing radiologists with advanced tools to detect subtle and elusive signs of cancer, particularly those leading to interval diagnoses, AI has the potential to significantly improve the accuracy and efficiency of breast cancer screening. This collaborative approach promises to usher in an era of more precise, personalized, and ultimately more effective cancer care, leading to better outcomes for countless patients.
The critical research was supported in part by significant contributions from the National Institutes of Health, the National Cancer Institute, the Agency for Healthcare Research and Quality, and Early Diagnostics Inc. The team of authors from UCLA included 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, whose collective efforts underscore the multidisciplinary nature of this vital scientific endeavor.
