PISDGAN: Perceive image structure and details for laryngeal image enhancement

Abstract

The introduction of optical techniques such as narrow-band imaging (NBI) during laryngeal endoscopic diagnosis can emphasize the superficial mucosal veins and increase the identification rate of lesions. To eliminate the limitation of endoscopic devices, in this paper, an image enhancement method that can perceive the information and details of the image structure (PISDGAN) is proposed. The method is based on the classical cycle consistency image translation architecture, and its generator is adjusted to a U-shaped residual structure. At the same time, a multiscale cross-layer adaptation module (CLA) is designed, which integrates features of different depths through 'feature transfer' to obtain abundant image information. A non-local feature denoising module (FD) is also introduced to enhance the quality of the translated image. In addition, the structure difference loss is proposed to constrain the image structure to ensure the integrity of the translated image. Based on the self-built white light and narrow band laryngeal image datasets, the experimental comparison of PISDGAN and other similar image translation methods shows that the proposed method not only effectively preserves the image structure, but also maintains good image quality while highlighting the details.