Breast Density Analysis Using Digital Breast Tomosynthesis

Abstract

Background: Breast density (BD) is a significant breast cancer risk factor, traditionally assessed from two-dimensional (2D) mammograms. Mammography has shifted to digital breast tomosynthesis (DBT). BD measures developed for traditional 2D mammograms will require translation to DBT images. Objectives: We evaluated an anutomated percentage of BD measure (PDa) using several methods as risk factors with DBT data: (1) normalized volumetric measure; (2) total dense volume (Dv); (3) 2D measure applied to DBT volume images (slices) and averaged (slice-mean); and (4) applied to synthetic 2D images. For the DBT volume analysis, measure were derived theoretically and evaluated. PDawas modeled as a function of compressed breast thickness. An alternative method for constructing synthetic 2D mammograms was evaluated using the PDavolume results as a basis. Other measures such as the mean and standard deviations of the pixel values were also investigated as risk factors. Methods: A case-control study (n = 426 pairs) design was investigated with matching on age, hormone replacement, imaging unit, and screening history. Conditional logistic regression modeling, controlling body mas index and ethnicity, was used to estimate odds ratios (ORs) for the various image measures with 95% confidence intervals provided parenthetically. ORs were estimated per standard deviation increment of the respective image measurement. Results: Volumetic PDawas significantly associated with breast cancer risk [OR = 1.43 (1.18, 1.72)] and produced an identical OR as the slice-mean measure. PDavalues from the slices (function of compressed breast thickness) were modeled as 2nddegree polynomial (concave-down). The maximum PDavalue occurred at approximately 0.41×(compressed breast thickness), and this position was similar across case-control groups; PDafrom this slice location was significant [OR = 1.47 (1.21, 1.78)]. PDaproduced significant findings when applied to the 2D synthetic DBT images: [OR = 1.44 (1.18, 1.75)]. The mean of the pixel values taken from either the volume or synthetic 2D images were similar and significant [ORs ~ 1.31 (1.09, 1.57)]. Dvdid not produce a significant finding. Discussion: This study produced several findings. The PDaalgorithm provided significant ORs when applied to the volume and synthetic 2D images. AS predicted by the derivations, the volumetric and slice-mean measures are equivalent and produced an identical OR. The maximum PDaslice position found, empirically agreed with the polynomial model prediction. Dvwas not significant until normalized by the breast volume. The volumetric analysis enabled the construction of an alternative standardized 2D synthetic image, where each pixel value represents the percentage of BD above its location. Conclusion: Several BD measurements derived from DBT images produced significant breast cancer associations, showing the validity of the automated techniques. Theoretical derivations supported by empirical analyses led to the construction of a new synthetic 2D standardized image derived from the volume measurements. Ancillary to the study objectives, these volumetric findings provide evidence for understanding the nature of the percentage of breast density measure applied to traditional 2D mammograms. The study will require replication in different populations. Notwithstanding the findings, the study design provides a template for investigating other possible risk measures such as texture from DBT.