Deep learning-based denoising for synchrotron radiation breast computed tomography images

Introduction:
Breast cancer is among the most prevalent oncological diseases in women. While mammography remains the gold standard for screening, synchrotron radiation phase-contrast breast computed tomography (SR-PCbCT) provides improved soft-tissue discrimination and enhanced visualization of fine structural details without tissue overlap. However, radiation dose remains a limitation: reducing the dose increases image noise, compromising image quality. Deep learning-based denoising is therefore investigated to mitigate these effects.

Methods:
A supervised U-Net model was developed for denoising monochromatic (32 keV) SR-PCbCT projections acquired at the SYRMEP beamline of the Elettra synchrotron (Trieste, Italy). Low-dose acquisitions (mean glandular dose of 5 mGy) were denoised using corresponding high-dose (25 mGy) data as supervision targets. A residual learning strategy was adopted, modeling low-dose projections as the sum of high-dose signal and noise. The network predicts noise, which is subtracted to obtain the denoised output. Two approaches were tested: an unconstrained method, in which the full predicted noise is removed, and a constrained method, in which a scaling factor controls the amount of subtracted noise. Phase retrieval was applied either before or after denoising to evaluate the influence of its ordering on the reconstructed output. Performance was assessed using image quality metrics (such as contrast, Signal-to-Noise ratio, and spatial resolution) and global fidelity metrics (such as Structural Similarity Index Metrics and Mean Squared Error).

Results:

Unconstrained denoising reduces noise but introduces over-smoothing, with loss of high-frequency content and degradation of image texture. Constrained denoising enables controlled behavior: low values of subtracted noise preserve structural detail, while higher values degrade texture and structural visibility. Phase retrieval after denoising better preserves image quality.

Conclusions:
Deep learning-based denoising improves SR-PCbCT image quality, but performance depends on denoising strength and phase retrieval order. Best results are obtained at intermediate levels of subtracted noise with phase retrieval applied after denoising.