Low-dose CT denoising via convolutional neural network with an observer loss function

Purpose: Convolutional neural network (CNN)-based denoising is an effective method for reducing complex computed tomography (CT) noise. However, the image blur induced by denoising processes is a major concern. The main source of image blur is the pixel-level loss (e.g., mean squared error [MSE] and mean absolute error [MAE]) used to train a CNN denoiser. To reduce the image blur, feature-level loss is utilized to train a CNN denoiser. A CNN denoiser trained using visual geometry group (VGG) loss can preserve the small structures, edges, and texture of the image.However, VGG loss, derived from an ImageNet-pretrained image classifier, is not optimal for training a CNN denoiser for CT images. ImageNet contains natural RGB images, so the features extracted by the ImageNet-pretrained model cannot represent the characteristics of CT images that are highly correlated with diagnosis. Furthermore, a CNN denoiser trained with VGG loss causes bias in CT number. Therefore, we propose to use a binary classification network trained using CT images as a feature extractor and newly define the feature-level loss as observer loss. Methods: As obtaining labeled CT images for training classification network is difficult, we create labels by inserting simulated lesions. We conduct two separate classification tasks, signal-known-exactly (SKE) and signal-known-statistically (SKS), and define the corresponding feature-level losses as SKE loss and SKS loss, respectively. We use SKE loss and SKS loss to train CNN denoiser. Results: Compared to pixel-level losses, a CNN denoiser trained using observer loss (i.e., SKE loss and SKS loss) is effective in preserving structure, edge, and texture. Observer loss also resolves the bias in CT number, which is a problem of VGG loss. Comparing observer losses using SKE and SKS tasks, SKS yields images having a more similar noise structure to reference images. Conclusions: Using observer loss for training CNN denoiser is effective to preserve structure, edge, and texture in denoised images and prevent the CT number bias. In particular, when using SKS loss, denoised images having a similar noise structure to reference images are generated.