Quantifying the Impact of Energy-Domain Translation In Simulated Digital Mammography
Abstract
Purpose
To evaluate the accuracy and physical fidelity of deep learning–based energy-domain translation in simulated mammography using the VICTRE virtual clinical trial framework, with emphasis on image quality and lesion preservation.
Methods
Projection mammograms were generated using the VICTRE pipeline at 24 and 40 kVp, including datasets with and without inserted lesions. A modified Pix2PixHD-based convolutional neural network was trained to translate high-energy (40 kVp) projections to low-energy (24 kVp) projections using log-transformed projection images. Model outputs were reconstructed and evaluated against ground-truth 24 kVp DICOM projections. Multiple network configurations were explored; results from the best-performing configuration (Step 4) are reported. Image quality was assessed using paired comparisons of peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), mean absolute percentage error (MAPE), normalized mean squared error (NMSE), and lesion contrast-to-noise ratio (CNR) calculated within a breast-region mask. Lesion-containing and non-lesion cases were analyzed separately.
Results
The Step 4 configuration demonstrated strong agreement with ground truth across all datasets (PSNR = 37.35 ± 2.21 dB, SSIM = 0.899 ± 0.020, MAPE = 2.03 ± 1.08%, NMSE = 0.0061 ± 0.0085). For mass lesions with density ρ = 1.06 g/cm³, no statistically significant differences in CNR metrics were observed relative to ground truth (paired tests, p > 0.05). Higher-density mass lesions CNR (ρ = 1.25 g/cm³) exhibited statistically significant improvements relative to ground truth across all evaluated model steps. Calcification CNR showed larger localized residuals but remained visually conspicuous due to their high inherent contrast.
Conclusion
Deep learning–based energy-domain translation can accurately reproduce low-energy mammographic projections from high-energy inputs within a controlled simulation environment. Performance was robust for mass lesions, while calcification translation warrants further investigation. These findings establish a validated baseline for energy-domain translation within VICTRE and support continued investigation into its role in advanced mammographic imaging workflows.