By Billy Candra 
A new study led by investigators at the UCLA Health Jonsson Comprehensive Cancer Center suggests that artificial intelligence (AI) could help detect interval breast cancers—those that develop between routine screenings—before they become more advanced and harder to treat. This could potentially lead to better screening practices, earlier treatment, and improved patient outcomes. The groundbreaking research, published in the prestigious Journal of the National Cancer Institute, provides compelling evidence for the integration of AI into current mammography protocols, particularly within the distinct screening landscape of the United States.
Understanding the Challenge of Interval Breast Cancers
Breast cancer remains one of the most common cancers among women globally, with significant efforts focused on early detection to improve survival rates. Routine mammography screenings have been instrumental in this regard, dramatically reducing breast cancer mortality by identifying tumors at an early, more treatable stage. However, a persistent challenge in breast cancer screening is the phenomenon of "interval cancers." These are cancers diagnosed in the period between a negative mammogram and the subsequent scheduled screening. While representing a minority of all breast cancers, typically 20-30% depending on screening intervals and population demographics, interval cancers are often more aggressive, grow faster, and present at a later stage, making them notoriously difficult to treat and often associated with a poorer prognosis compared to screen-detected cancers.
The reasons for interval cancers are multifaceted. They can arise from rapidly growing tumors that emerge after a clear mammogram (true interval cancers), or they can be cancers that were present at the time of screening but were either missed by the interpreting radiologist (missed reading error), had very subtle signs that were not recognized (minimal signs), or were genuinely invisible on the mammogram (occult cancers). It is this latter category—cancers visible on mammograms but missed by human interpretation due to their subtlety—where AI holds particular promise. The human eye, despite extensive training and experience, can sometimes overlook faint or ambiguous signs, especially amidst dense breast tissue or the sheer volume of images radiologists must review. The UCLA study specifically targeted these "mammographically-visible" types of interval cancers, aiming to leverage AI’s analytical capabilities to augment human perception.
The UCLA Health Study: Methodology and Key Findings
The retrospective study, led by Dr. Tiffany Yu, assistant professor of Radiology at the David Geffen School of Medicine at UCLA and first author, and Dr. Hannah Milch, assistant professor of Radiology at the David Geffen School of Medicine and senior author, analyzed an extensive dataset comprising nearly 185,000 past mammograms performed between 2010 and 2019. This substantial historical data included both Digital Mammography (DM), often referred to as 2D mammography, and Digital Breast Tomosynthesis (DBT), commonly known as 3D mammography. From this large cohort, the research team meticulously identified 148 cases where a woman was subsequently diagnosed with interval breast cancer.
To understand why these cancers were not detected earlier, a panel of expert radiologists meticulously reviewed each of the 148 interval cancer cases. They adapted a comprehensive European classification system to categorize the reasons for non-detection at initial screening. This system meticulously breaks down interval cancers into several categories:
- Missed reading error: The cancer was visible on the mammogram but overlooked by the radiologist.
- Minimal signs-actionable: Subtle signs of cancer were present, which, in retrospect, should have prompted further action.
- Minimal signs-non-actionable: Extremely subtle signs were present, arguably below the threshold of reliable human detection at the time.
- True interval cancer: The cancer developed rapidly after a truly negative screening.
- Occult: The cancer was genuinely invisible on the mammogram at the time of screening.
- Missed due to a technical error: Issues with image acquisition or processing led to the cancer being missed.
Following this detailed human review, the researchers applied a commercially available AI software, Transpara, to the initial screening mammograms of these 148 interval cancer cases. The objective was to determine if the AI tool could identify the subtle signs of cancer that were missed by radiologists during their initial interpretation, or at least flag these images as suspicious. The AI software generated a risk score for each mammogram, ranging from 1 to 10, indicating the likelihood of cancer. A score of 8 or higher was predetermined as the threshold for flagging a mammogram as potentially concerning.
The study’s findings were compelling. Researchers estimate that the strategic incorporation of AI into screening protocols could lead to a significant reduction in the number of interval breast cancers, potentially by as much as 30%. This reduction is primarily attributed to AI’s ability to identify "mammographically-visible" types of interval cancers earlier. Dr. Tiffany Yu 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. 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." Earlier detection often translates to smaller tumors, less lymph node involvement, and thus, less invasive surgeries, fewer rounds of chemotherapy or radiation, and ultimately, a higher chance of long-term survival and improved quality of life.
Bridging the Gap: U.S. vs. European Screening Paradigms
While similar research exploring the utility of AI in breast cancer detection has been conducted in Europe, the UCLA study stands out as one of the first to specifically investigate its application within the distinct context of the United States. This distinction is crucial due to fundamental differences in screening practices and technological adoption between the two regions.
In the U.S., the landscape of breast cancer screening is characterized by a high adoption rate of Digital Breast Tomosynthesis (DBT), or 3D mammography. DBT provides a series of thin-slice images of the breast, which can be reconstructed into a 3D view, allowing radiologists to see through overlapping breast tissue that can obscure cancers in traditional 2D mammography. This technology is widely considered superior for detecting cancers, especially in women with dense breasts, and has become the standard of care in many U.S. facilities. Furthermore, U.S. screening guidelines typically recommend annual mammograms for women starting at age 40 or 50, depending on risk factors and organizational guidelines.
In contrast, many European breast cancer screening programs historically have relied more heavily on conventional Digital Mammography (DM), or 2D mammography. While 2D mammography is effective, it lacks the depth perception of DBT. Moreover, European screening intervals are generally longer, with many countries adopting biennial (every two years) or even triennial (every three years) screening schedules. These differences in technology and frequency mean that AI tools developed and validated in European settings might not directly translate or perform optimally within the U.S. healthcare system without specific adaptation and re-evaluation. The UCLA study, by utilizing a dataset predominantly featuring DBT and reflecting typical U.S. screening frequencies, provides invaluable data tailored to the American context, making its findings particularly relevant for U.S. radiologists and policymakers.
AI’s Promise and Perils: Accuracy and Limitations
Despite the highly encouraging results regarding AI’s potential to reduce interval cancers, the study also provided a nuanced perspective on the current limitations and "inaccuracies" of AI in real-world scenarios. Dr. Hannah Milch, the study’s senior author, candidly addressed these challenges: "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."
One striking example of this dual nature of AI performance related to "occult" cancers – those truly invisible on mammography. The AI tool, Transpara, surprisingly flagged 69% of the screening mammograms that subsequently developed occult cancers as suspicious. This indicates a sensitivity of the AI to subtle patterns that might precede the visual manifestation of cancer, or perhaps to other features that correlate with higher risk, even if not directly cancerous. However, when researchers delved deeper to assess the AI’s ability to pinpoint the exact location of the cancer within these flagged images, its performance was significantly less accurate. The AI only marked the actual cancer location 22% of the time. This disparity highlights a crucial challenge: AI can identify a risk or suspicion in an image, but its precision in localization, especially for nascent or truly occult lesions, is still a work in progress.
This finding underscores that while AI can serve as a powerful "second set of eyes" or a risk stratification tool, it is not yet a standalone diagnostic solution. It may flag an area as suspicious, prompting a radiologist to scrutinize it more closely, even if the AI’s precise localization is off. This necessitates a careful integration strategy where human expertise remains paramount, augmented by AI, rather than replaced by it. The implications for workflow are significant: how do radiologists handle cases where AI flags areas as suspicious that are not visible to the human eye, especially when the AI’s pinpointing accuracy is limited? These are complex questions that require further investigation through larger, prospective studies.
Expert Perspectives and Clinical Implications
The insights from the UCLA study offer a glimpse into a future where AI could fundamentally alter the workflow of radiologists and enhance the precision of breast cancer screening. For radiologists, AI tools like Transpara could act as intelligent assistants, sifting through vast numbers of images and highlighting areas of concern that might otherwise be overlooked due to fatigue, time constraints, or the sheer volume of cases. This could potentially reduce the cognitive burden on radiologists, allowing them to focus their expertise on the most challenging cases.
Dr. Yu’s concluding remarks encapsulate this vision: "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 means that AI could help catch those cancers that are visible but missed, leaving primarily those that genuinely develop rapidly between screenings or are truly undetectable. "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 potential benefits extend beyond individual patient outcomes. From a public health perspective, a reduction in interval cancers could lead to a decrease in the overall burden of advanced breast cancer, potentially lowering healthcare costs associated with more intensive treatments. Patient advocacy groups have long championed initiatives that enhance early detection, recognizing its direct correlation with improved survival rates and quality of life. The prospect of AI making screening even more effective would undoubtedly be welcomed by these organizations, fostering greater confidence in screening programs.
Broader Context: The Role of AI in Diagnostic Imaging
The UCLA study is part of a broader global trend of integrating artificial intelligence into medical imaging. AI algorithms, particularly deep learning models, have shown impressive capabilities in various diagnostic areas, from detecting subtle lesions in X-rays and CT scans to analyzing pathology slides and ophthalmic images. In radiology, AI promises to address several pressing issues: improving diagnostic accuracy, reducing interpretation time, combating radiologist burnout, and standardizing image interpretation across different practitioners.
However, the journey to widespread AI adoption in healthcare is not without hurdles. Beyond the technical limitations highlighted in the UCLA study, there are significant regulatory, ethical, and practical challenges. Regulatory bodies like the FDA are developing frameworks for approving AI algorithms, focusing on their safety, efficacy, and transparency. Ethical considerations include potential biases in AI algorithms (e.g., if trained on unrepresentative datasets), data privacy, and accountability when diagnostic errors occur. Practically, integrating AI into existing clinical workflows requires significant infrastructure investment, robust IT support, and extensive training for medical professionals.
Future Directions and Unanswered Questions
The retrospective nature of the UCLA study, while providing valuable insights, naturally leads to the call for larger, prospective studies. These future studies would involve integrating AI into the real-time screening process, observing how radiologists interact with AI alerts, and measuring the actual impact on interval cancer rates and patient outcomes in a live clinical setting. Key questions that prospective studies need to address include:
- How do radiologists effectively incorporate AI flags into their decision-making process, especially when AI indicates suspicion in areas invisible to the human eye?
- What is the optimal threshold for AI flagging to maximize cancer detection while minimizing unnecessary recalls and biopsies (false positives)?
- How does AI perform across diverse patient populations, including different breast densities, ethnicities, and age groups?
- What are the cost-effectiveness implications of integrating AI into screening workflows?
- What training is required for radiologists and technologists to effectively utilize AI tools?
Ultimately, the goal is not for AI to replace human radiologists, but to augment their capabilities, creating a synergistic partnership that leverages the strengths of both. The UCLA Health Jonsson Comprehensive Cancer Center study provides a powerful foundation, suggesting that this partnership could usher in a new era of breast cancer screening—one where technology empowers clinicians to detect the hardest-to-find cancers earlier, leading to more lives saved and improved patient well-being.
The work was supported in part by crucial funding from the National Institutes of Health, the National Cancer Institute, the Agency for Healthcare Research and Quality, and Early Diagnostics Inc., highlighting the collaborative effort required to advance such significant medical research. Other contributing 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.
