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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Convolutional neural network (CNN) denoising reduces computed tomography (CT) noise but can cause image blur.
• Pixel-level losses (e.g., MSE, MAE) used in CNN training are a primary source of this blur.
• Existing feature-level losses, like VGG loss, are suboptimal for CT due to reliance on natural image datasets and potential CT number bias.

Purpose of the Study:

• To develop a novel feature-level loss, termed observer loss, for training CNN denoisers on CT images.
• To mitigate image blur and preserve structural details lost with pixel-level losses.
• To address the CT number bias introduced by VGG loss in CT image denoising.

Main Methods:

• Created simulated lesions in CT images to generate labels for training a binary classification network.
• Defined two types of observer loss: signal-known-exactly (SKE) loss and signal-known-statistically (SKS) loss.
• Trained CNN denoisers using SKE and SKS observer losses.

Main Results:

• CNN denoisers trained with observer loss preserved image structure, edges, and textures better than those trained with pixel-level losses.
• Observer loss successfully resolved the CT number bias issue associated with VGG loss.
• SKS loss resulted in denoised images with noise structures more closely resembling reference images compared to SKE loss.

Conclusions:

• Observer loss is an effective method for training CNN denoisers, enhancing structural preservation and preventing CT number bias in CT images.
• SKS loss, a specific type of observer loss, yields denoised images with noise characteristics similar to the original, high-quality reference images.