Low-dose CT denoising via CNN with an observer loss function

Convolutional neural network (CNN) based low-dose CT denoising is effective to deal with complex CT noise. However, CNN denoiser with pixel-level loss functions (e.g., mean-squared-error (MSE) and mean-absoluteerror (MAE)) often produces image blur. To overcome this limitation, perceptual loss function (e.g., VGG loss) is adapted to train CNN denoiser. CNN denoiser with VGG loss preserves structural details in denoised images better. However, because VGG network is trained with natural RGB images, which have different image properties from CT images, extracted features would not be related to diagnosis. Also, CNN denoiser with VGG loss introduces a bias of CT number in denoised images. In this work, we propose observer loss to train CNN denoiser. Observer network (i.e., feature extractor in observer loss) is trained with CT images to classify lesion-present and lesion-absent cases. We conduct two binary classification tasks, signal-known-exactly (SKE) and signal-known-statistically (SKS) tasks. Because it is hard to obtain labeled CT images, we insert simulated lesions. CNN denoiser with observer loss preserves small structures and edges in denoised images without introducing bias in CT number.