UC Davis

Objective: To generate low-dose Breast CT (LD-bCT) images from a cone-beam Breast CT (bCT) system and denoise them to achieve a comparable image quality to their standard-dose (SD) counterparts.Methods: LD-bCT phantom images were generated from standard-dose bCT (SD-bCT) phantom datasets for specific pixel binning and gain modes using four parameters unique to each flat-panel detector. The four parameters determined were F, the mean signal-variance relationship, σ2e, the variance of electronic noise, Kq, the autocorrelation kernel of quantum noise, and Ke, the autocorrelation kernel of electronic noise. A noise injection equation incorporating the four parameters was formulated and applied to SD phantom datasets. The noise power spectrum (NPS) was used as a similarity metric to validate the noise injection procedure for LD-bCT simulation. The mean squared error (MSE) and noise variance quantified the difference between the curves. After the verification, LD-bCT patient data was generated at 75 and 50% of the SD level. Transfer learning with a pretrained network called the Residual Encoder-Decoder CNN (RED-CNN) was utilized for the LD-bCT denoising task since it was previously trained on Low-Dose CT data. A variation of the Hyperbane hyperparameter tuning algorithm was performed to identify the optimal hyperparameters. The loss metric was chosen as MSE, and the evaluation metrics were selected as the root mean squared error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM). The model's overall performance was validated using adversarial inputs, the testing set, and another test set that appreciated for all 18 patient datasets. Results: The variance of electronic noise was calculated as 82.747ADU2, the F factor as 0.00125, Ke as a 2D delta function, and Kq as a 7x7 correlation matrix. The NPS curves had good agreement with MSE values of 2.22e-14, 3.56e-14, and 4.82e-13 at 90, 60 and 45mA, respectively. After applying transfer learning using a RED-CNN and deciding optimal parameters for the denoising task, the model was trained for 150 epochs and obtained a validation MSE loss of 5.11e-5. During the testing stage, the model obtained an RMSE of 0.0097, PSNR of 41.0308, and SSIM of 0.9673. Conclusion: The LD-bCT simulation supplies an accurate noise modeling technique for generating LD-bCT images with excellent NPS agreement with its corresponding SD-bCT. When LD-bCT patient data was fed to an extension of the pre-trained RED-CNN, the images are found to be effectively denoised without comprising on detail and while preserving anatomical structures. Clinically, this research finds applications in LD-bCT patient acquisitions and SD-bCT enhancement. Keywords: Computed Tomography, Denoising CNNs, Low-Dose Image Denoising