Purpose: To simulate low dose CT images from clinical dose acquisitions in order to create datasets for testing image analysis or processing algorithms.
Methods: A dataset of 2020 paired 2D slices from 101 clinical dose (400 mA; 80 kVp) and low dose (200 mA; 80 kVp) 320 slice non-contrast brain CT scans, reconstructed using 3D iterative reconstruction, was obtained retrospectively. Rigid registration was performed between the clinical dose and low dose images to correct for patient movement. Two neural networks were separately trained on this data for, first, denoising and then, generating the necessary noise expected at lower dose. The denoising network has a U-net architecture with residual layers, which estimates the average of the distribution, denoising the images. The noise generating network is a generative adversarial network (GAN), of which the generator is a U-net with residual layers and the discriminator is a modified version of the VGG16 model. The generator has to output realistic noise patterns for low dose CT images, in which the spatial correlation and magnitude depend on the content, location, and intensity of the image. As the real input image for the discriminator, the noise pattern obtained by subtracting the denoised image from the low dose image was used.
Results: The denoising network reduces the noise effectively, with a reduction in the pixel variance in homogeneous areas by a factor of nine. The generator of the GAN produces realistic noise patterns of low dose images, which have spatial, content and intensity dependent correlation and magnitude, and it is able to generate realistic noise even across the edge of the skull.
Conclusion: Using the denoising and generator networks enables generating multiple low dose images from a single clinical dose image with realistic noise patterns, opening the possibility to create datasets for testing purposes.
Funding Support, Disclosures, and Conflict of Interest: This work was partially funded by Canon Medical.