Application of neural networks to denoising of CT images of lungs

One of the most challenging problems in the field of digital image processing is image denoising. When processing medical images, it is of particular relevance to improve the perceived quality of images, while preserving the diagnostically relevant information. This paper investigates the capacity of a neural network framework for medical image denoising. Specifically, the performance of the proposed image denoising method is evaluated on a database of computed tomography images of lungs using various image quality metrics, such as peak signal-to-noise ratio and mean squared error. Image denoising relies on block segmentation of noisy and low-pass filtered images to generate the input and the target data for the neural network training. This paper investigates how the choice of block size, network architecture, and the training method affect the denoising performance when image is degraded with additive Gaussian noise. The paper proposes the use of Kohonen's self-organizing maps for segmentation of feature space and the use of multiple, finely tuned multi-layer perceptrons to achieve an improved denoising performance.