The impact of noise texture on a deep learning denoising model for high-resolution cardiac EID-CT

A convolutional neural network was trained for denoising high-resolution (HR) energy-integrating detector (EID) cardiac CT images. To supervise the training, noise-only images were estimated from patient images using two different methods: Subtraction of low and high iterative reconstruction (IR) strength reconstructions; subtraction of adjacent image slices with the same IR strength. After training, the models were applied to independent HR cardiac EID-CT images and compared to HR photon-counting detector (PCD)-CT images. Models trained with either noise-image generation method reduced noise in HR EID-CT images by 74-79% to achieve comparable noise to HR PCD-CT images. However, the model trained on subtraction of low and high IR images resulted in undesirable salt-and-pepper noise texture and CT number bias. The model trained on adjacent slice subtraction images had more natural texture and no significant bias. Both models preserved spatial resolution of the inputs and did not alter anatomical structures.