Toward denoising of 3D CT scans with few data

Abstract

We propose a new method for denoising of 3D CT scans with few data. Like any other form of imaging data, CT scans are susceptible to noise and artifacts. Noise in CT scan images is not only stochastic, but can be frequency dependent and introduced by the measuring device itself or by signal processing algorithms. Unfortunately, most state-of-the-art Deep Learning methods are mainly focusing on denoising random noise only. Therefore, we propose a new method for denoising 3D CT scans, which is based on a 3D AutoEncoder, a GAN, and self-supervised learning. Our method works not only on random noise but also the frequency dependent noise. It exploits the fact that, in a CT scan image, an object’s features are similar between different regions and easy to be encoded. In contrast, the features of structured noise are very different from the object, while the random noise is pixel-wise independent. Our method can be trained on few data, with or without ground truth, and is computationally inexpensive. Our experiments show that our method outperforms other methods in terms of several metrics, and outperforms most state-of-the-art methods in terms of computational efficiency.