|
Cone-beam breast CT (bCT) provides volumetric images of the uncompressed breast but present higher noise than 2D mammography. Deep Learning (DL) denoising with supervised training has shown successful CBCT noise reduction but requires matched low-dose and high-dose images. Self-supervised training removes that requirement but often assume locally independent noise. This work studies the impact of bCT noise correlation on self-supervised denoising methods. The self-supervised training strategies included two blind spot methods – Noise2Self, enforcing local similarity with independent image noise; and Noise2Sim, enforcing image similarity in presence of correlated noise – and two Noisier2Noise approaches: i) noise injection in the image domain; and, ii) noise injection in projection domain with a model of noise correlation. Self-supervised training was performed on bCT images generated from 150 voxelized models with a high-fidelity forward projector, including models of the x-ray spectrum, polychromatic attenuation, and detector signal and noise propagation. Denoised images were assessed with respect to high-dose references and supervised denoising, using RMSE, SSIM, and noise power spectrum (NPS). Noise2Sim and Noisier2Noise with noise injection in the projection domain showed good performance in presence of correlated noise, achieving RMSE of 0.21 and 0.18 (SSIM of 0.9 and 0.94), respectively, compared to RMSE of 0.17 (SSIM of 0.93) for supervised training. The independent noise assumption in Noise2Self and Noisier2Noise with image domain noise injection resulted in significantly diminished performance, yielding RMSE of 0.23 and 0.37 (SSIM of 0.86 and 0.84). The NPS measurements revealed a shift towards low frequency components for Noise2Sim, arising from blurring of tissue boundaries and residual image transfer induced by the masking of dissimilar regions in the loss function. Noisier2Noise showed a frequency distribution of noise closer to the high-dose reference. Such performance was slightly degraded for non-matched noise injection models inducing shorter correlation kernels than the nominal detector noise correlation, but models inducing longer correlation showed negligible impact in the denoising results. Self-supervised denoising in presence of correlated noise was proved feasible. Among the evaluated models, Noisier2Noise strategies with projection domain noise injection showed denoising performance comparable to supervised training and noise spectral distribution comparable to high-dose bCT.
|
