Imposing noise correlation fidelity on digital breast tomosynthesis restoration through deep learning techniques

Task-based assessment of digital mammography microcalcification detection with deep learning denoising algorithms using in silico and physical phantom studies

“…In both, we used an existing PyTorch deeplearning denoising model provided to us by its developers. 7,10 In the first set of experiments, the denoising model was trained (by the algorithm creators) using clinical data and applied to images of a physical breast phantom with embedded microcalcifications acquired on the Hologic Selenia Dimensions DM system. In the second set of experiments, a cohort of digital anthropomorphic breast phantoms was generated with embedded microcalcification clusters, and mammograms were simulated with Monte Carlo x-ray transport software.…”

“…Details of its implementation and training are given by Shan et al, 7 and the source code is available from USP-LAVI GitHub repository. 17,18 To briefly summarize the authors developments, a HResNet network (modification of a ResNet architecture 19 optimized for better modeling of the noise distribution in low-dose DM) has been investigated with seven different loss functions. Specifically, the authors experimented with MSE, MAE, SSIM, 8 and PL [compares similarity between two images in a high-level feature space of the VGG-16 network 20 ] loss functions PL1 through PL4, with a numerical index indicating whether the PL was computed on early or later layers in the VGG-16 for four different feature spaces.…”

“…Training For the study with Monte Carlo generated data, we employed an improved version of the denoising algorithm by the same group (Laboratory of Computer Vision, University of São Paulo, Brazil), presented at the IWBI 2022 conference. 10 Similar to the original work, our implementation used a combination of two loss functions: one to impose fidelity in the noise correlations (2D power spectrum or PS-2D) and one that optimizes visual perception (perception loss or PL) in VGG-16 network. The PS-2D component was defined as the mean-normalized-absolute erro between the 2D power spectrum of the prediction (e.g., model output image) and the target (ground-truth high-dose image); whereas the PL component compares similarity between two images in a high-level feature space in the early or later layers of the VGG-16 (thus dubbed PL1 through PL4) and is equivalent to the MSE but defined in a feature space instead of a pixel space.…”

“…Results suggested that one PL loss function (PL4) was able to achieve similar noise levels of the full-dose image when applied to reduced dose images, while achieving smaller bias than the other loss functions tested. In a follow up study, Vimieiro et al 10 enhanced the loss function to impose fidelity of the image noise correlation between full-dose and low-dose mammograms. In another study, Gao et al 11 trained a DL-CNN using a loss function consisting of a weighted combination of MSE and adversarial loss to improve microcalcification conspicuity in breast tomosynthesis images.…”

“…Results suggested that one PL loss function (PL4) was able to achieve similar noise levels of the full-dose image when applied to reduced dose images, while achieving smaller bias than the other loss functions tested. In a follow up study, Vimieiro et al 10 . enhanced the loss function to impose fidelity of the image noise correlation between full-dose and low-dose mammograms.…”