An emerging artificial intelligence algorithm, developed to estimate volumetric breast density from 3D-reconstructed digital breast tomosynthesis images, could potentially facilitate individual risk assessments for breast cancer.
An artificial intelligence (AI) algorithm developed to estimate volumetric breast density (VBD) from three-dimensional (3D)-reconstructed digital breast tomosynthesis (DBT) images showed potential in a study presented at the Radiological Society of North America (RSNA) 2022 Annual Meeting. Moreover, researchers found that the algorithm’s VBD estimation (which does not require raw image data) and body mass index (BMI) were associated with breast cancer risk.
“Such a tool can enable large retrospective epidemiological and personalized risk assessments of breast density with DBT,” said Vinayak Ahluwalia, B.S.E., an M.D. candidate at the Perelman School of Medicine at the University of Pennsylvania, Philadelphia, and colleagues.
While the use of DBT is increasing in breast cancer screening, current methods estimate VBD from two-dimensional (2D) images and often require raw DBT data, which is rarely stored by clinical centers. In this study, Ahluwalia and colleagues developed a computational tool based on deep learning for the assessment of VBD using only 3D-reconstructed DBT images. They also analyzed the ability of the tool to perform individualized breast cancer risk assessment.
The study included data from 1,080 negative DBT screening exams obtained between September 2011 and November 2016 from the Hospital of the University of Pennsylvania. Both 2D raw and 3D-reconstructed DBT images were available in craniocaudal and mediolateral oblique views, with 7,850 views in total. The mean age of the participants was 57 years, the mean BMI was 28.7 kg/m2 and the racial makeup of the cohort was 41.2 percent White, 54.2 percent Black and 4.6 percent other individuals. The researchers generated corresponding 3D reference-standard tissue segmentations from previously validated software that employed both 3D reconstructed slices and raw 2D DBT data.
Ahluwalia and colleagues trained a 3D deep learning model that used U-Net architecture to create a three-label image segmentation task for background, dense tissue, and fatty tissue. The dataset was randomly divided into training (70 percent), validation (15 percent) and test (15 percent) sets. The labels were evaluated overall and separately with the weighted Dice score (DSC), with 0 signifying no overlap and 1 signifying perfect overlap.
When the AI algorithm was evaluated on an independent testing set, it achieved unweighted DSC of 78 percent and weighted DSC of 56 percent. The algorithm accurately segmented the three labels of background, fatty tissue, and dense tissue with respective DSCs of 94 percent, 89 percent, and 49 percent. The AI algorithm performed similarly among racial groups, according to the study.
The researchers also evaluated whether the VBD estimated from the algorithm and BMI were associated with breast cancer risk in an independent case-control set of 193 patients who developed breast cancer and 714 control patients from the same hospital system. The results showed that VBD (odds ratio of 1.27) and BMI (odds ratio of 1.37) were statistically significant factors in individualized breast cancer prediction.
“We have created an AI algorithm for estimating VBD from 3D-reconstructed DBT images that does not require the use of raw image data and could aid in predicting an individual’s risk of breast cancer,” noted the study authors.