Perceptual Transformer Loss for the Neural Denoising of Ultra-Low Dose CT

Neural image denoising has proven a powerful approach for both generic and medical image enhancement. In recent years, fully convolutional architectures, followed by GANs, and more recently visual transformers have consecutively defined the state of the art in the field. A critical component to the success of the neural approach is the loss under which the model is trained. In particular, the rich features generated by CNN-based perceptual loss methods have demonstrated strength in leading to a denoised image quality consistent with human perception. In this work, we present a natural evolution of the CNN-based loss wherein transformer-based features are used to define improved perceptual losses. The proposed perceptual transformer losses (PETEL) are validated qualitatively and quantitatively against CNN-based perceptual loss, as well as L2 and L1 losses for the training of state-of-the-art network architectures for an ultra-low dose Computed Tomography (CT) denoising task. A dataset of 37 real ultra-low dose - normal dose CT scan pairs is utilized for this purpose. The obtained results suggest that the proposed PETEL are promising losses for the challenging neural denoising of ultra-low dose CT images.