Delving Into Density
By Beth W. Orenstein
Vol. 23 No. 4 P. 16
AI tools help accurately and consistently classify breast density on mammograms.
Breasts contain different types of tissue that can be seen on a mammogram: glandular, connective, and fat. Dense breasts have relatively high amounts of glandular tissue and fibrous connective tissue and little fatty breast tissue. The ACR has developed BI-RADS to group different types of breast density. BI-RADS Fifth Edition, published in 2013, describes breast density as an “assessment of the volume of attenuating tissue in the breast” that is classified as the following four categories:
• A) almost entirely fatty (about 10% of women);
• B) scattered areas of dense glandular tissue and fibrous connective tissue (about 40% of women);
• C) heterogeneously dense (about 40% of women); and
• D) extremely dense (about 10% of women).
Mammograms are harder to read in women with dense breast tissue. “Dense breast tissue can mask or hide a cancer on a mammogram,” says Constance Lehman, MD, PhD, director of breast imaging and codirector of the Avon Comprehensive Breast Evaluation Center at Massachusetts General Hospital in Boston.
“Women need to know whether they have dense breasts because it puts them at slightly higher risk for a future breast cancer event,” Lehman says. “Women need to know whether they have dense breasts so that they can choose to undergo supplementary tests to improve cancer detection.”
Federal legislation passed in 2019 requires all mammography providers to inform patients and physicians about their breast density after they undergo a mammogram and explain the importance of that information. The problem is that determining breast density on a mammogram is highly subjective.
“It’s a metric that is important, but, unfortunately, there is wide human variation in accurately assessing a woman’s density,” Lehman says. Even when using BIRADS, two or more radiologists can interpret the same mammogram differently, or the same radiologist can make different assessments on the same person, she adds.
Subjectivity is one reason researchers are turning to AI to solve this problem. A study published in March 2022 in Radiology: Artificial Intelligence by researchers in Italy found that an AI tool can accurately and consistently classify breast density on mammograms. The AI tool they developed for breast density classification, TRACE4BDensity, was based on a sophisticated type of AI known as deep learning with convolutional neural networks. It is capable of discerning subtle patterns in images beyond the capabilities of the human eye, says corresponding author Christian Salvatore, MSc, PhD, CEO of DeepTrace Technologies and a researcher at the University School for Advanced Studies IUSS Pavia. The tool is available upon request at firstname.lastname@example.org.
The tool was trained using the breast density category that was determined by the majority of seven board-certified radiologists who independently assessed 760 mediolateral oblique images in 380 women with a mean age of 57 at one center. Three radiologists whose breast density assessment was closest to the majority of the initial seven performed external validation of the model. Their validation was performed on a data set of 384 mediolateral oblique images in 197 women with a mean age of 56-plus years obtained from a second center. The AI tool the Italian researchers developed and externally validated had an 89.3% accuracy for nondense vs dense breast classification and substantial agreement of 90.4% with radiologist readings, the authors wrote.
Salvatore says the next step for the research team is to apply their AI technology to screening breast studies to evaluate its impact on women with dense breasts. However, he believes, “using this system, radiologists will be able to assign women a breast density category in a reliable and reproducible way, overcoming the variability of human readers, which is currently a substantial limitation in clinical practice.”
FDA-Approved AI Tools
Lehman, a professor of radiology at Harvard Medical School, also believes that AI tools will help reduce variation and improve the quality of density assessments. “That’s the goal of the Italian study, and others, including those that we have done,” she says. Many practices have incorporated AI in their mammography screening using tools that are on the market and FDA approved for that purpose.
To date, the FDA has approved eight AI tools for addressing breast density. “They are trained in different ways to assess density on mammograms,” Lehman says. Some look at the volume of dense breast tissue in the entire breast, and some look for patterns and texture. Others look for whether the density is uniformly spaced across the breast or concentrated in some areas.
Among the available tools are Hologic’s Quantra, Volpara Health’s TruDensity, and iCAD Inc’s PowerLook Density Assessment. Hologic’s first iteration of AI-powered Quantra software debuted in 2008, Volpara’s TruDensity has been in clinical use since 2010, and iCAD launched its first breast density assessment technology in 2015 under the name iReveal. Quantra breast density assessment software is designed to provide an objective method of assessing a patient’s breast density, says Jennifer Meade, president of the breast and skeletal health division at Hologic. Hologic’s software uses a machine learning algorithm that analyzes each patient’s pattern and texture to provide an unbiased breast density assessment. The information includes the Quantra Density Category, which is consistent with the four breast composition categories of the ACR BI-RADS Atlas, Fifth Edition, Meade says.
Quantra software focuses on pattern and texture for breast density assessment, Meade says. “Many other software solutions measure breast density volumetrically, calculating the total volume of fatty vs dense breast tissue by volume,” she says. However, with the Fifth Edition of BI-RADS, “the industry best practice for calculating breast density has been updated to recognize that volumetric calculations are not always sufficient for calculating breast density,” Meade says. “This is because pattern and texture also play a significant role.”
For example, she says, when a woman has dense tissue uniformly spaced across her breast, many breast cancers can potentially be identified in the pockets of fatty tissue. However, Meade explains, in a breast where the same volume of dense breast tissue is condensed into a singular area, breast cancers can more easily hide behind the concentrated tissue.
“The industry has recognized that a breast with uniformly spaced breast density may qualify for a lower BI-RADS score, compared with a breast that has concentrated breast density, even though both have the same total volume of dense tissue,” Meade says.
Commenting on the AI study by the Italian researchers, Meade says, “It’s always promising to see research affirming the viability of AI-powered breast density assessment software to accurately identify breast density within the BI-RADS scoring system.” If breast density assessment software becomes ubiquitous throughout the industry, patients will benefit from the improved continuity of scoring and the removal of subjective interradiologist visual assessment, she adds.
The hope is that when breast density score information from this type of software is paired with case and lesion scores from AI-powered breast cancer detection software, radiologists will have more valuable information to confidently make recommendations for supplemental imaging needs. “This could be particularly important as facilities develop protocols and best practices for patients who have been deemed high risk and who have dense breasts,” Meade says.
Hologic’s Quantra breast density assessment software calculates a patient’s breast density score at the point of care on the acquisition workstation, allowing technologists to review scores immediately, Meade says. “After previewing on the acquisition workstation, the breast density score transfers with the corresponding image to the radiologist for reading. This information ultimately translates into the patient report,” she explains.
The score previews immediately on the acquisition workstation, allowing technologists to instantly flag patients with extremely dense breast tissue with a radiologist. “Depending on a facility’s breast density protocols, this may prompt the radiologist to expedite reading and potentially request supplemental diagnostic imaging within the same day,” Meade says. “This scenario can be especially beneficial for patients who do not need to get called back for a separate appointment for supplemental imaging.”
Ralph Highnam, PhD, chief science and innovation officer of Volpara Health, says more than 13.5 million women have had their breast density analyzed with TruDensity since its debut in 2010, making it one of the original AI tools for measuring breast density. Volpara TruDensityuses a combination of X-ray physics and AI to generate objective volumetric measurements of breast composition. TruDensity determines density from the raw data in the mammogram. “Because that information does not vary, TruDensity is very reproducible and offers a valuable decision support tool to inform the radiologist’s own assessment of density,” Highnam says.
TruDensity automatically assesses the volumetric breast density percentage of each mammogram on a continuous scale. The typical volumetric breast density range is 2% to 35%. This differentiates each woman on a continuum of density. These Volpara Density Grade scores correlate with the BI-RADS Fifth Edition, which is clear that breast composition assessment is about the volume of attenuating, or dense, tissue, Highnam explains.
In fact, the Fifth Edition indicates that beyond being just an assessment of the volume of fibroglandular (dense) tissue in the breast, radiologists should also consider increasing the density category assessment if the fibroglandular tissue in a breast “is sufficiently dense to obscure small masses.” The key to this statement is, of course, judging what “sufficiently dense” means. To do this, “one needs to understand breast composition in a volumetric sense,” he says.
To build trust that AI improves breast imaging, AI products must be appropriately designed and validated through rigorous processes, to ensure performance and accuracy, Highnam says. “Currently,” he adds, “there are no widely accepted standards for design or validation of AI today, which is why many tools come to market having been developed and tested on only a few hundred cases. Looking to the future, more research is needed to incorporate characteristics such as pattern and texture into density assessment to ensure consistent, explainable AI.”
Highnam says Volpara Health has taken steps to establish best practices for the creation of its AI. “This includes using large, diverse data sets for training and testing to ensure better generalization; participating in independent, third-party prospective validation studies; and a physics-based approach that is fundamentally explainable,” he says.
To ensure that its algorithm is robust and generalizable, Volpara Health used more than 70 data sets with more than 50 million images and cases from facilities around the world. More than 36 tests are performed for every software release to ensure that it works with every FDA-approved mammography system and that it can resolve for implants, pacemakers, image quality issues, and other “noise” that clinicians see in the real world but generally are not included in “clean” training data sets, Highnam says.
“In addition,” he says, “we perform tests to compare how it performs on digital 2D mammography vs 3D tomosynthesis, how it scores the same woman when the exam is taken on different mammography systems, and how it performs compared to radiologists.” The accuracy of Volpara TruDensity has been validated by many independent studies, showing high correlation to breast MRI, which is considered the source of ground truth in breast density. “We are proud to have evaluated the breast density of women screened in the 10-year DENSE Trial,” Highnam says. The DENSE Trial demonstrated that offering MRI to women with very dense glandular breast tissue and normal results on screening mammography reduced the number of interval cancers by 50%.
Stacey Stevens, president and CEO of iCAD, says the latest version of its Power-Look Density Assessment is the world’s first and only multivendor, deep learning, automated breast density assessment algorithm using synthetic images generated from 3D mammography. The algorithm analyzes the structure, texture, and dispersion of the fibroglandular tissue and provides a personalized score that can help standardize stratification in density-based breast cancer screening.
“It works by identifying the patient’s anatomy, segmenting the breast, then measuring adipose and fibroglandular tissue and its dispersion, to determine the density category in alignment to the BI-RADS Fifth Edition lexicon,” Stevens says. “Results can be integrated into common risk models that incorporate density and automated mammography reports through leading medical imaging system/RIS software products.”
While all women can benefit from a personalized approach to breast screening protocols, Lehman says breast density needs to be a large part of that equation. “Decisions are being made based on their breast density,” she says, “So, if we are giving them accurate, consistent information, we can guide them down the clinical care pathway most appropriately.”
Accuracy is important, Lehman says, because a woman with fatty breast tissue who is told she has dense breast tissue may unnecessarily undergo additional screenings and possible biopsies. Even more importantly, a woman with dense breast tissue who doesn’t know it could falsely believe she is of average risk and wouldn’t benefit from supplemental testing. If AI can improve the consistency and the information regarding a woman’s dense breast tissue, as it appears to, Lehman believes it would help significantly.
— Beth W. Orenstein of Northampton, Pennsylvania, is a freelance medical writer and regular contributor to Radiology Today.