Pulmonary nodule image super-resolution using multi-scale deep residual channel attention network with joint optimization

Abstract

High-resolution medical images can help doctors to find early lesions and provide assistance and support for the diagnosis and treatment of diseases. Super-resolution can obtain a single high-resolution image from a given low-resolution image. It can be divided into three major means roughly: the interpolation-based methods, the reconstruction-based methods, and the learning-based methods. The interpolation-based methods rely on the smoothness assumptions and cannot restore fine textures. The reconstruction-based methods need to find the image degradation model and the optimal blur kernels. In fact, blur kernels are complicated and unknown. Kernel mismatch will fail to produce good results (e.g., over-sharpening or over-smoothing). The learning-based methods pay more attention to the understanding of the image content and structure, and they can establish a mapping function between the high-resolution images and the low-resolution images, which attract the attention of researchers. Deep learning has the strong ability of nonlinear mapping. Therefore, it has been widely used in super-resolution in recent years. In the paper, we propose a multi-scale deep residual channel attention network which consists of six components: joint input of low-resolution image and edge, shallow feature extraction, deep feature extraction, channel attention, high-resolution image reconstruction, and total loss. Edges are the first-order high-frequency details which are very important to super-resolution. The joint input of low-resolution images and edges enhances useful information. The multi-scale deep residual channel attention module can not only acquire structural features but also capture features of different scales and hierarchies. It can also obtain relationships among channel features. In addition, the joint guidance of perceptual loss, content loss, and edge loss is used to improve the visual quality and preserve the spatial structure and high-frequency details of low-resolution images. Experiments have been conducted on the pulmonary nodule image dataset, and the results demonstrate that the proposed method can yield better performance by comparing with the state-of-the-art methods.