Author: Alexander Bookbinder, Matthew Tivnan, Xiangyi Wu, Wei Zhao 👨🔬
Affiliation: Stony Brook Medicine, Massachusetts General Hospital 🌍
Purpose: To investigate and benchmark a system-adaptive diffusion-based digital breast tomosynthesis (DBT) denoising model for a direct-indirect dual-layer flat panel detector (DI-DLFPD) with a k-edge filter at the tube port. Effective denoising models mitigate tradeoffs of dual-energy DI-DLFPD setups, including filter material constraining dose allocation between direct and indirect detectors, while preserving benefits like eliminating patient motion artifacts. Improving image signal-to-noise ratio is particularly important for DBT, with relatively lower dose-per-projection compared to digital mammography, yielding correspondingly higher quantum noise. This is especially true for contrast enhanced DBT (CEDBT), where tissue structural noise is removed and x-ray quantum noise increased during LE and HE image weighted subtraction.
Methods: The noise power spectrum (NPS) and modulation transfer function (MTF) were characterized for DLFPD with tungsten target and 100-micron silver tube port-located k-edge filter. Breast phantoms, healthy and lesioned, were generated using VICTRE, and were used to generate noise-free DBT projections via ray tracing with ASTRA. Images for both detector layers were blurred with corresponding MTF, and noise was added incrementally, causing NPS agreement with system performance at given dose levels. A diffusion deep-learning model was trained on the reversed noise addition, using pseudo-inverse diffusion techniques previously demonstrated for low-dose CT. Images were denoised in the spatial projection domain. Performance was validated using physical phantoms at various dose levels.
Results: The model trained using this process showed enhanced ability to recover and enhance fine details, including microcalcifications. System-adaptive models were more efficient to train, converged more quickly and with lower mean squared error than Gaussian noise-based models.
Conclusion: We demonstrate the utility of adapting diffusion denoising models to novel DLFPD-based CEDBT systems by designing the forward process to reflect system-specific NPS and MTF, yielding faster and more accurate denoising. Pseudo-inverse diffusion techniques show promise for DBT dose reduction and image quality enhancement.