Comparison of Digital Breast Tomosynthesis and Mammography-based Radiomics for Breast Cancer Risk Assessment: Case-Control Study

Abstract

Purpose To compare the performance of volumetric radiomic parenchymal pattern analysis from three-dimensional (3D) digital breast tomosynthesis (DBT) images with that of two-dimensional (2D) digital mammography (DM) and 2D sections from DBT in assessing breast cancer risk relative to breast density measurements. Materials and Methods This was a retrospective matched case-control study among individuals who underwent concurrent DM and DBT screening from March 2011 through December 2014. The Cancer Phenomics Toolkit was used to calculate radiomic features from craniocaudal and mediolateral oblique views in all study patients, matched on race and age, for various experimental settings, including image resolution and window size. For each image type, conditional logistic regression evaluated the association of radiomic features, along with age, body mass index (BMI), and area percent density (PD) (from the Laboratory for Individualized Breast Radiodensity Assessment software), with breast cancer, using the C statistic as the measure of model predictive ability. Model fit was compared via likelihood ratio tests. Results The study included 924 female patients (median age, 61 years [IQR: 51-69 years]), with 187 cases and 737 controls. Volumetric features from 3D reconstructed DBT scans had, on average, higher C statistics across all experimental conditions. Among models using only radiomic features, C statistics were highest for models using features from 3D images (mean C statistic: 0.68, P < .001); models using features from 2D image types resulted in lower mean C statistics (0.60 to 0.65). A baseline model using age, BMI, and area PD had a C statistic of 0.60. The effect of higher image resolution and smaller window size were not substantial, supporting the use of less computationally intensive processing. Conclusion Fully automated 3D parenchymal analysis from DBT improved breast cancer risk estimation beyond markers derived from area breast density and 2D images. Keywords: Mammography, Tomosynthesis, Breast, Volume Analysis Supplemental material is available for this article. © RSNA, 2025.

Keywords: Breast; Mammography; Tomosynthesis; Volume Analysis.

Graphical abstract

Figure 1:

Exclusion flowchart for study patients from March 2011 to December 2014. BI-RADS = Breast Imaging Reporting and Data System, BMI = body mass index, PD = percent density.

Figure 2:

Study design. (A) Data acquisition. (B) Modeling and analysis. (Figure is simplified, does not include radiomic analysis at different window sizes and resolutions.) DBT = digital breast tomosynthesis, DM = digital mammography, 3D = three-dimensional.

Figure 3:

Schematic of fully automated three-dimensional lattice-based feature extraction with window size (W), resolution resampling size (D), and number of sections (N). The example patient image is a left mediolateral oblique view from a digital breast tomosynthesis image in a 41-year-old female patient who was a control.

Figure 4:

Graph shows C statistics of six different conditional logistic regression models of cancer risk, for all window sizes and resampling resolutions, for the six image types analyzed. Vertical bars indicate the means. BI-RADS = Breast Imaging Reporting and Data System, BMI = body mass index, DBT = digital breast tomosynthesis, DM = digital mammography, PC = principal component, PD = percent density, 3D = three-dimensional.

Figure 5:

Graph shows C statistics versus resampling resolutions, using the radiomics-only model (feature principal components 1–10 and no clinical or demographic covariates), for window size of 12.8 mm, for all image types. DBT = digital breast tomosynthesis, DM = digital mammography, 3D = three-dimensional.

Figure 6:

Graphs show C statistics versus resampling resolutions, using the radiomics-only model (feature principal components 1–10 and no clinical or demographic covariates), for all window sizes, for image types (A) raw DM, (B) DBT central projection, processed, and (C) DBT reconstructed (3D). The lack of obvious trend versus resampling resolution is typical over all window sizes. DBT = digital breast tomosynthesis, DM = digital mammography, 3D = three-dimensional.

Figure 7:

Graph shows P values for likelihood ratio tests of three conditional logistic regression models of cancer risk, each versus the parallel model without radiomic information, for all window sizes and resampling resolutions, for the six image types analyzed. Vertical bars indicate the means. A lower P value indicates more reason to believe that the model with radiomic information is a better fit than the parallel model without (low P value indicates evidence against the null hypothesis that the models with and without radiomics are equivalent). BI-RADS = Breast Imaging Reporting and Data System, BMI = body mass index, DBT = digital breast tomosynthesis, DM = digital mammography, PC = principal component, PD = percent density, 3D = three-dimensional.