By Lester Litchfield

Breast cancer is the most common cancer in women, accounting for about 30% of all new female cancers annually. With one in eight women in the U.S. developing breast cancer, the healthcare industry must adopt a nuanced approach to screening that considers individual medical histories and risk factors.

Early detection of breast cancer significantly improves patient outcomes as it often leads to less invasive treatments. However, predicting who is at the highest risk is complex.

Risk factors like breast tissue density, genetic predispositions and lifestyle factors all contribute to the complexity of determining who is at the highest risk, underscoring the need for personalized screening strategies that go beyond age-based guidelines.

Mammography: Strengths and limitations
Mammography remains the gold standard for early detection, with over 41 million procedures performed annually. However, mammograms miss about 20% of existing breast cancers due to image quality variability and the difficulty of detecting cancer in dense breast tissue.

Breast density refers to the relative amounts of different kinds of fibrous, glandular, and fatty tissue, and is both a robust and independent risk factor for breast cancer and a factor impacting the sensitivity of mammography. On a mammogram, both dense tissue and tumors appear white, which can obscure potential cancers. As breast density increases, the sensitivity of mammograms decreases, making it harder to detect cancerous growths early. Dense tissue also increases your risk for developing breast cancer.

According to dense breast-info.org, some cancers will not show on a mammogram in women with dense breasts. Additional tests, such as ultrasound, molecular breast imaging, contrast enhancement, and MRI after a mammogram, may find these cancers. It is important to note, however, that women should not skip their mammogram, as some cancers are only seen on a mammogram, even in dense breasts.

Ordinarily, radiologists visually and subjectively assess a patient’s breast density to place them in one of four categories according to definitions set out in the American College of Radiology’s Breast Imaging Reporting and Data System (BI-RADS®). Alternatively, AI can automatically assess the volume of dense tissue present in a patient’s breast on a continuous scale to help ensure a more precise identification and consistent application of screening protocols. Research has shown that radiologists using only subjective, visual assessment assign density categories to patients inconsistently, and the variability between how different radiologists categorize density is quite high. Breast density can be objectively and quantitatively assessed as a percentage: the amount of dense tissue divided by the total breast tissue. Women with smaller breast volumes often have a higher density percentage, which further complicates cancer detection.

This is not a health challenge for a small population. Having dense breast tissue is common. Nearly half of US women over the age of 40 have dense breasts. Additionally, breast density can vary across different demographics. For instance, Asian women generally have higher breast density. This variability drove the creation of different scales for measuring breast density in regions like Japan, where using the standard U.S. scale would categorize most women on the higher end of the density spectrum.

The role of AI in advancing breast cancer detection
AI technology can address the limitations of traditional breast cancer screening methods, particularly in cases involving dense breast tissue.

AI algorithms, trained on large and diverse datasets, can enhance image analysis by identifying subtle patterns and anomalies that are difficult for the human eye to see, particularly in dense breast tissue.

Moreover, AI can assist in personalized risk assessments by analyzing a patient’s medical history, genetic predispositions, and other risk factors. This allows for more tailored screening recommendations and potentially identifies high-risk individuals earlier than traditional guidelines. For example, AI can help triage women with denser breast tissue for supplemental imaging like MRI or ultrasound, improving the likelihood of early detection.

AI's ability to integrate with existing systems and adapt to different imaging technologies further enhances its utility across diverse populations. By reducing false negatives and improving the accuracy of early detection, AI is a powerful tool in the fight against breast cancer.

Global advancements in personalized breast cancer screening
Regulatory bodies and medical societies are promoting personalized screening approaches. For example, the FDA has updated breast density reporting requirements, recognizing dense tissue as a risk factor, and the American College of Radiology (ACR) recommends breast cancer risk assessments for all women by age 25 to identify high-risk individuals early.

Internationally, the European Society of Breast Imaging (EUSOBI) has recommended that countries mandate notifying women of their breast density. This recommendation is driven by research from the DENSE trial in the Netherlands, which demonstrated that supplemental imaging for women with dense breasts significantly improves cancer detection rates.

A unique aspect of the DENSE trial was the use of AI technology to objectively identify the women with the densest breast tissue (category D or higher). These women were then triaged to receive MRI screening, a strategy that effectively managed the limited MRI capacity in the Netherlands by focusing on those at the highest risk.

The movement to address breast density in screening is not limited to Europe. Similar guidelines and practices are being considered and implemented in countries like Australia and New Zealand. The Royal Australian and New Zealand College of Radiologists (RANZCR) is also focusing on integrating AI and other technologies into breast cancer screening protocols, particularly for women with dense breasts.

Additionally, the Find It Early Act, introduced in Congress, aims to ensure that all women, regardless of insurance coverage, have access to breast cancer screening and diagnostic tests, enhancing early detection efforts and addressing disparities in access to care.

Trends and challenges in AI for mammography
Recent studies indicate that commercial AI algorithms can perform as well or above the level of human radiologists. The benefit of AI to breast cancer detection is particularly seen in clinical settings where radiology is experiencing significant backlogs and staff shortages.

In the U.S., where a single reader reads mammograms, radiologists are being tasked with reading increasing volumes without letting quality slip and missing cancers. Preliminary results of the AI-STREAM prospective study found that adding AI into radiologists' workflow increased the cancer detection rates of both General Radiologists and Breast Radiologists by 25% and 14%, respectively. The Breast Radiologists with AI had the highest cancer detection rates and lowest recall rates, providing an important quality safety net to radiologists in an increasingly high-pressure environment.

Another clinical setting facing growing pressures are European screening services that routinely use double reading; where two radiologists independently review each case, with a tiebreaker process if they disagree. With the global shortage of radiologists, maintaining this labor intensive protocol is leading to delays and strains on the screening service. . Recently, a major prospective study— the ScreenTrustCAD trial, conducted in Scandinavia— demonstrated AI’s potential to elevate this radiologist shortage. AI was used to replace one of the human readers, and the results showed that AI was non-inferior and, in some cases, even superior to having two human readers.

The success of these trials underscores the importance of conducting prospective studies under real-world conditions, where AI is used on actual patients within ongoing clinical practice. This approach contrasts with retrospective trials, which often utilize research datasets that may not represent the diversity of X-ray machines, population characteristics, and other variables encountered in clinical settings.

AI has also shown promise in reducing the rate of interval cancers (cancers diagnosed between regular screening intervals) and triaging mammograms by prioritizing those with suspected findings.

Despite these advances, the adoption of AI in mammography is still in the early stages, with many radiologists yet to integrate these tools into their practice. Issues such as costs, ethical concerns, and patient privacy need to be addressed to facilitate broader adoption.

The importance of diverse and robust training datasets
AI systems require large and diverse datasets for training to ensure accuracy across different populations. The diversity of equipment, patient populations, and even the processing algorithms used by different vendors can significantly impact AI performance.

For example, if an AI system is trained primarily on images from one type of machine, it might not perform as well when analyzing images from another. An example of this comes from a study from Scotland, where a software update to image preprocessing significantly increased the cancellation rate of AI-flagged cases because the AI had not been trained on images processed with the new software.

Companies like Volpara, which utilize unprocessed images rather than fully processed ones, have an advantage in this regard. By training AI on unprocessed images, they can avoid the issues associated with varying processing algorithms.

Evaluating AI solutions for breast imaging
Imaging centers should consider several factors to ensure they implement the most effective and reliable technology.

As previously discussed, the size and diversity of the AI dataset used to train algorithms are paramount. A larger and more varied dataset helps improve the algorithm's performance and ensures it can accurately detect cancer across different demographics and breast tissue types.

Third-party validation studies are another essential consideration. Relying solely on studies funded or published by the company that developed the AI tool may not provide a complete picture of its efficacy. In fact, in a recent overview of 100 CE-marked AI products from 54 vendors, 64 of the 100 products reviewed had no peer-reviewed evidence of their efficacy.

Explainable AI features, which provide transparency into how the AI makes decisions, are also important, as trust in AI is often a barrier to its adoption in clinical settings.

As always, compliance support is critical in a heavily regulated field like healthcare. AI vendors must prioritize compliance with all relevant regulations, including the new Mammography Quality Standards Act (MQSA) requirements that took effect effect 10September 2024.

Finally, selecting an AI vendor for breast imaging committed to continuous innovation in personalized cancer screening is essential to prevent technological obsolescence. A vendor focused on long-term program success, not just algorithms, ensures you deliver cutting-edge care while fostering a partnership invested in improving patient outcomes.

About the author: Lester Litchfield is the head of data science for Volpara Health, a Lunit company.