AI could help improve early detection of interval breast cancers

The Silent Threat: Understanding Interval Breast Cancers

Breast cancer remains one of the most prevalent cancers globally, with millions of diagnoses each year. While advancements in screening technologies like mammography have significantly improved early detection rates and patient survival, a persistent challenge in breast imaging is the occurrence of "interval breast cancers." These are malignant tumors that manifest clinically between scheduled screening mammograms, typically within 12 to 24 months of a normal or benign screening result. Unlike screen-detected cancers, interval cancers often present with more aggressive biological features, are diagnosed at a later stage, and consequently carry a poorer prognosis. They represent a critical area of concern for both radiologists and patients, accounting for a significant percentage of all breast cancer diagnoses in screened populations, with estimates ranging from 20% to 30% in some studies.

The difficulty in detecting these cancers lies in their often subtle or non-existent signs on prior mammograms, or their rapid growth dynamics that allow them to develop and become symptomatic within the screening interval. For patients, an interval cancer diagnosis can be particularly distressing, raising questions about the efficacy of their previous screening and leading to heightened anxiety. For clinicians, identifying the underlying reasons for non-detection is crucial for improving screening protocols and diagnostic accuracy. This background underscores the urgent need for innovative solutions to bridge the detection gap that current screening methods occasionally leave open.

UCLA’s Groundbreaking AI Study: A New Frontier in Detection

Investigators at the UCLA Health Jonsson Comprehensive Cancer Center have embarked on a pivotal study, published in the esteemed Journal of the National Cancer Institute, exploring the potential of artificial intelligence to significantly mitigate the challenge of interval breast cancers. The research team, led by Dr. Tiffany Yu, assistant professor of Radiology at the David Geffen School of Medicine at UCLA and first author, along with senior author Dr. Hannah Milch, also an assistant professor of Radiology, delved into retrospective data to assess if AI could identify "mammographically-visible" types of interval cancers earlier by flagging them at the time of initial screening.

The study specifically focused on two critical categories of interval cancers that are technically visible on mammograms but often elude human detection: tumors that are present but simply missed by radiologists during their review (missed reading error), and those with very subtle signs that are either too faint or arguably below the threshold of human perception at the time of screening (minimal signs-actionable or non-actionable). By leveraging AI, the researchers hypothesized that a computational tool, unburdened by human fatigue or the inherent limitations of visual acuity, could act as an invaluable "second set of eyes." The findings suggest a remarkable potential: researchers estimate that incorporating AI into screening could help reduce the number of interval breast cancers by an impressive 30%.

"This finding is important because these interval cancer types could be caught earlier when the cancer is easier to treat," emphasized Dr. Yu. "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 statement encapsulates the profound impact such a technology could have on patient care, shifting the paradigm from reactive diagnosis to proactive, earlier intervention.

Methodological Rigor: Data, AI Tool, and Classification

To conduct their robust analysis, the UCLA team performed a retrospective study, meticulously analyzing data from nearly 185,000 past mammograms collected between 2010 and 2019. This extensive dataset encompassed both digital mammography (DM), often referred to as 2D mammography, and digital breast tomosynthesis (DBT), commonly known as 3D mammography. The inclusion of both modalities is crucial, reflecting the transition in screening technologies over the past decade and allowing for a comprehensive evaluation of AI’s performance across different imaging types.

From this vast pool of data, the researchers specifically identified and scrutinized 148 cases where a woman had been diagnosed with interval breast cancer. These cases formed the core of their investigation. Following the identification of these interval cancers, a panel of experienced radiologists meticulously reviewed the initial screening mammograms to determine why the cancer was not detected earlier. To standardize this assessment, the new study adapted a well-established European classification system to categorize the interval cancers into distinct types:

• 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, could have prompted further investigation.
• Minimal signs-non-actionable: Very subtle signs were present, but their significance was too low to warrant action at the time.
• True interval cancer: The cancer was genuinely not visible on the prior mammogram, indicating rapid growth or development within the screening interval.
• Occult: The cancer was truly invisible on the mammogram, confirming no mammographic manifestation.
• Missed due to a technical error: Issues with image acquisition or processing led to non-detection.

Once the interval cancers were classified, the researchers applied a commercially available AI software called Transpara to the initial screening mammograms performed before the cancer diagnosis. Transpara is designed to assist radiologists by identifying suspicious areas. The tool assigned a cancer risk score to each mammogram, ranging from 1 to 10. A score of 8 or higher was designated as a "flag" by the AI, indicating a potentially concerning finding that warranted closer scrutiny. The objective was to determine if the AI could detect the subtle signs of cancer that were missed by human radiologists during their initial screenings, or at least flag them as suspicious, thereby prompting a second look.

Bridging the Atlantic: US vs. European Screening Paradigms

A significant aspect of the UCLA study’s contribution is its focus on the U.S. screening landscape, which presents key differences from European practices where similar AI research has been more prevalent. Understanding these distinctions is crucial for appreciating the unique implications of this study for American healthcare.

In the United States, the predominant mammography technology used for screening is digital breast tomosynthesis (DBT), or 3D mammography. DBT offers several advantages over traditional 2D digital mammography (DM) by providing multiple thin-slice images of the breast, which helps to reduce tissue overlap – a common limitation of 2D mammograms that can obscure cancers or create false positives. Furthermore, U.S. screening guidelines typically recommend annual mammograms for women starting at age 40 or 50, depending on individual risk factors and organizational guidelines. This more frequent screening interval is intended to catch cancers earlier.

Conversely, European screening programs have historically relied more heavily on 2D digital mammography (DM). While 2D mammography is effective, it lacks the multi-slice advantage of DBT. Moreover, European guidelines often recommend longer screening intervals, typically every two to three years. These differences in technology and screening frequency have direct implications for the types of interval cancers observed and the potential for AI to intervene. A longer screening interval, for instance, might naturally lead to a higher proportion of "true interval cancers" that develop rapidly between screenings.

The UCLA study, by predominantly analyzing data that includes DBT and reflecting annual screening practices, provides invaluable insights directly relevant to the American healthcare system. This makes the findings particularly significant for guiding future AI integration strategies and policy decisions within the U.S., as it accounts for the specific challenges and opportunities presented by current American screening modalities and frequencies.

Key Findings and the Nuances of AI Performance

The application of Transpara AI to the retrospective mammograms yielded compelling, albeit nuanced, results. The AI demonstrated a significant capacity to identify interval cancers that had been missed by human radiologists. This capability underpins the researchers’ estimate of a 30% reduction in interval breast cancers if AI were routinely incorporated into screening workflows. Such a reduction would translate into thousands of lives saved and improved quality of life for countless patients through earlier, less aggressive treatment options.

However, the study also provided a realistic assessment of AI’s current limitations and "inaccuracy issues" that warrant further exploration in real-world settings. Dr. Hannah Milch, the senior author, highlighted a critical observation: "For example, despite being invisible on mammography, the AI tool still flagged 69% of the screening mammograms that had occult cancers. However, when we looked at the specific areas on the images that the AI marked as suspicious, the AI did not do as good of a job and only marked the actual cancer 22% of the time."

This particular finding reveals a dual nature of current AI capabilities. On one hand, the AI’s ability to flag a high percentage of occult cancers (those genuinely invisible to the human eye on mammograms) suggests it might be detecting subtle, imperceptible patterns or changes that even advanced human perception cannot discern. This could potentially open entirely new avenues for detecting cancers that were previously considered impossible to identify mammographically. On the other hand, the low accuracy (22%) in pinpointing the exact location of these occult cancers is a significant practical challenge. If AI flags a broad area as suspicious without precise localization, it could lead to unnecessary follow-up imaging, biopsies, increased patient anxiety, and a substantial burden on healthcare resources. This emphasizes that while AI can detect "something," its current precision in localization for the most challenging cases still requires refinement.

The study’s detailed classification of interval cancers further illuminated where AI’s strengths and weaknesses lie. AI performed particularly well in identifying cases categorized as "missed reading errors" and "minimal signs-actionable," precisely the types of cancers that are mammographically visible but overlooked. This reinforces the idea that AI excels as a "second reader" for overt but subtle findings. For "true interval cancers" and "occult cancers," where the signs are genuinely absent or extremely faint, the AI’s performance transitions from precise detection to a more generalized "flagging" of potential concern, highlighting areas for further research and technological improvement.

Implications for Clinical Practice and Patient Outcomes

The potential implications of integrating AI into breast cancer screening are far-reaching, promising a paradigm shift in how radiologists operate and how patients experience their healthcare journey.

For Radiologists: AI is not envisioned as a replacement for human radiologists but rather as a powerful augmentative tool. It could serve as an intelligent assistant, helping to reduce the cognitive load on radiologists, especially given the increasing volume and complexity of mammograms. By pre-screening images or highlighting suspicious areas, AI could allow radiologists to focus their expertise on the most challenging cases, potentially reducing burnout and improving overall diagnostic accuracy. The study’s findings suggest AI could be particularly beneficial for less experienced radiologists or in high-volume screening settings where fatigue might contribute to missed readings. However, it also introduces new challenges: radiologists will need training on how to interpret AI flags, especially when the AI identifies something not visible to the human eye. This could lead to new diagnostic pathways, potentially involving other imaging modalities like ultrasound or MRI, even in the absence of a clear mammographic finding.

For Patients: The most profound impact would be on patient outcomes. A 30% reduction in interval cancers means more cancers are caught at an earlier, more treatable stage. This could translate directly into:

• Less aggressive treatment: Earlier detection often allows for lumpectomies instead of mastectomies, and potentially less extensive chemotherapy or radiation, preserving quality of life.
• Improved survival rates: Early-stage breast cancer has significantly higher five-year survival rates compared to late-stage diagnoses.
• Reduced anxiety: For many women, the period between screenings is fraught with anxiety, especially if they have risk factors. Knowing that an AI is also reviewing their mammograms could provide an added layer of reassurance.
• More equitable care: AI could potentially help standardize screening quality across different healthcare settings, reducing disparities in detection rates.

The prospect of AI shifting the balance of interval breast cancers towards "mostly true interval cancers" (where the cancer genuinely develops too rapidly to be seen on the prior screening) is particularly exciting. As Dr. Yu noted, "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."

Challenges and Future Directions

While the UCLA study illuminates a promising path forward, it also candidly highlights the significant challenges that must be addressed before AI can be widely integrated into clinical practice.

Need for Prospective Studies: The current study was retrospective, meaning it looked back at past data. The next critical step is to conduct large-scale prospective studies. These studies would involve integrating AI into real-time screening workflows and observing its performance and impact in a live clinical setting. Such studies are essential to understand:

• How radiologists actually use AI in practice.
• The workflow implications and efficiency gains or losses.
• The rate of false positives and false negatives generated by AI, and their impact on patient anxiety and healthcare costs.
• How to manage cases where AI flags suspicious areas that are not visible to the human eye, especially when AI’s localization accuracy is still limited.

Regulatory Hurdles: The integration of AI into medical diagnostics is subject to rigorous regulatory oversight. AI tools must demonstrate consistent safety, efficacy, and clinical utility to gain approval from bodies like the U.S. Food and Drug Administration (FDA). The "black box" nature of some AI algorithms, where the reasoning behind a decision is not transparent, can also pose a challenge for regulatory approval and physician acceptance.

Ethical Considerations: The use of AI in healthcare raises several ethical questions, including:

• Accountability: Who is responsible if an AI makes an error that leads to a missed diagnosis or an incorrect one?
• Bias: AI algorithms are trained on vast datasets. If these datasets are not diverse and representative of the population, the AI could perpetuate or even amplify existing health disparities.
• Patient trust: How will patients react to the idea of an algorithm participating in their diagnosis?

Integration into Workflow and Training: Seamless integration of AI into existing radiological workflows will be crucial for adoption. This includes developing user-friendly interfaces, ensuring interoperability with existing imaging systems, and providing comprehensive training for radiologists and technologists on how to effectively utilize and trust AI tools.

A Collaborative Future for AI in Cancer Detection

The UCLA study, supported in part by significant funding from the National Institutes of Health, the National Cancer Institute, the Agency for Healthcare Research and Quality, and Early Diagnostics Inc., represents a crucial step forward in the application of AI for breast cancer detection. The collaborative effort involving a broad team of UCLA authors, 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, underscores the multidisciplinary nature of this research.

The findings resonate with a broader trend in medical imaging, where AI is increasingly being explored for its potential to enhance diagnostic accuracy across various modalities and disease states. From detecting subtle signs of retinopathy in ophthalmology to identifying lung nodules in radiology, AI is proving to be a powerful tool. However, as the UCLA study thoughtfully demonstrates, AI is not a panacea. Its true power lies in its ability to augment human expertise, providing a valuable "second opinion" that can catch what might otherwise be missed.

The journey from promising research to widespread clinical adoption is often long and complex, fraught with technical, regulatory, and ethical considerations. Yet, the vision of a future where AI significantly reduces the burden of interval breast cancers – saving lives and improving outcomes for countless women – serves as a powerful motivator for continued innovation and rigorous scientific inquiry. The UCLA Health Jonsson Comprehensive Cancer Center’s work marks a significant milestone on this vital path.