Deep learning reconstruction of digital breast tomosynthesis images for accurate breast density and patient-specific radiation dose estimation

TL;DR

This paper addresses the challenge of deriving true breast density and patient-specific radiation dose from digital breast tomosynthesis (DBT), a task hampered by limited vertical resolution. It introduces DBToR, a data-driven reconstruction framework that unrolls a proximal–dual optimization into a neural network and conditions reconstruction on the breast-thickness prior, enabling quantification of fibroglandular tissue from DBT projections. Trained on virtual phantoms and patient-derived breast CT phantoms with Poisson-noised projections, DBToR achieves density estimates within about $\pm 3\%$ and dose estimates within about $\pm 20\%$ without systematic bias, outperforming MLTR, LPD, and U‑Net baselines. The work demonstrates robust performance under noise and lays groundwork for true patient-specific dosimetry and density-based risk stratification, with potential to inform personalized screening strategies and dose registries. A key limitation is the slice-by-slice approach; future work will extend to full 3D volumes and incorporate polychromatic spectra and scatter for clinical applicability.

Abstract

The two-dimensional nature of mammography makes estimation of the overall breast density challenging, and estimation of the true patient-specific radiation dose impossible. Digital breast tomosynthesis (DBT), a pseudo-3D technique, is now commonly used in breast cancer screening and diagnostics. Still, the severely limited 3rd dimension information in DBT has not been used, until now, to estimate the true breast density or the patient-specific dose. This study proposes a reconstruction algorithm for DBT based on deep learning specifically optimized for these tasks. The algorithm, which we name DBToR, is based on unrolling a proximal-dual optimization method. The proximal operators are replaced with convolutional neural networks and prior knowledge is included in the model. This extends previous work on a deep learning-based reconstruction model by providing both the primal and the dual blocks with breast thickness information, which is available in DBT. Training and testing of the model were performed using virtual patient phantoms from two different sources. Reconstruction performance, and accuracy in estimation of breast density and radiation dose, were estimated, showing high accuracy (density <+/-3%; dose <+/-20%) without bias, significantly improving on the current state-of-the-art. This work also lays the groundwork for developing a deep learning-based reconstruction algorithm for the task of image interpretation by radiologists.

Figures (11)

• Figure 1: Network architecture of DBToR. Dual blocks (green) are on the upper row and primal blocks (blue) are in the bottom row. The blocks have the same architecture, elaborated in the first blocks. $m$: the breast thickness mask, $h_0$: initial dual vector, $f_0$: initial primal vector, $g$: sinogram data, $\text{Out}$: final reconstruction
• Figure 2: 2D coronal breast phantom containing skin (darkest gray), adipose tissue (dark gray), fibroglandular tissue (light gray), and Cooper's ligaments (black).
• Figure 3: (a) Coronal slice of a breast CT image, (b) the same image classified into skin (white), adipose (dark gray) and fibroglandular (light gray) tissue voxels, and (c) the classified deformed image with the technique described in Section \ref{['sec:compression']}.
• Figure 4: Imaging geometry implemented in the Monte Carlo simulation: the x-ray source is placed at 70cm from the detector, a 3mm thick polyethylene terephthalate (PET) compression paddle was simulated and a large water cuboid was included to take into account the patient-body backscatter. The x-ray field irradiated the breast model at different angles (from -24° to 24°). The center of rotation is placed at 65 cm from the x-ray source. Drawing is not to scale and rotation of the detector is not shown.
• Figure 5: Example of U-Net reconstruction artifact