Deep Residual Learning for Iodine Estimation in Digital Breast Tomosynthesis.

Abstract

Objective: Dual-energy contrast-enhanced digital breast tomosynthesis (DECE-DBT) could provide the quantification of iodine contrast with a pseudo-three-dimensional lesion reconstruction, outperforming CE two-dimensional mammography, and hence be a cost-effective alternative to DCE magnetic resonance for breast cancer imaging. However, limited-angle artifacts in DBT reduce its quantitative accuracy. In this work, we propose a deep learning (DL)-based method for DECE-DBT that performs artifact-robust material decomposition for iodine concentration estimation.

Methods: Evaluation was performed over 869 digital compressed breast phantoms containing heterogeneous lesions with varying iodine concentrations. The DL network consists of residual blocks, each containing a three-layer CNN followed by a U-Net, and predicts voxel-wise fractions of adipose-blood, fibroglandular-blood and iodine-blood mixtures.

Results: Optimization of the number of residual blocks showed that the 2-block model was optimal for the available dataset, generating hallucinations in under 2% of the tested cases. The network performance strongly depends on lesion size: the average dice similarity coefficient was 0.82 for lesions larger than 0.7 cm in diameter. Therefore, early-stage breast tumors (≤ 2 cm) fall mostly within the size range where the network performs reliably.

Conclusion: Aimed to shorten the path in making DECE-DBT a functional modality, these findings demonstrate the early potential of our DL-based approach for accurate quantification of iodine concentrations.