Cross Domain Deep Learning for Noise Removal from LDCT Images

In this paper, an unsupervised cross domain deep learning framework for noise removal from LDCT (Low Dose Computed Tomography) images is proposed. To tackle with the problem that paired LDCT image training data for noisy images and the corresponding clean versions are rarely available in real scenarios, the presented model is trained in an unsupervised manner with unpaired training images. Relying on the auto-encoder architecture, our framework aims at learning two deep generators/decoders for producing clean and noisy images, respectively, by learning disentangling feature representations. In our generators, our deep residual blocks and the spatial and channel attention modules with dilated convolutions are used for learning multi-scale image features. To train our image generators, unsupervised adversarial domain adaption is used to learn robust and invariant image representations. For each unpaired training images, by feeding the encoded image content features from the noisy training image to the clean image generator, we aim at generating the corresponding clean image. On the other hand, by feeding the encoded features, fused by the encoded content features from the clean training image and the encoded noise features from the noisy training image, to the noisy image generator, we aim at synthesizing the noisy image. To ensure the stability of this cross domain transformation, the cycle consistency loss is also included to keep the backward transformation from the synthesized images to their respective original domains. Extensive experiments on LDCT image denoising have shown the feasibility of the proposed framework.