Generalized degradation-based adversarial learning for unsupervised super-resolution of endomicroscopy images

In recent years, probe-based confocal laser endomicroscopy (pCLE) has become an emerging optical biopsy method for in situ imaging and diagnosis, which aids in the accurate early diagnosis of diseases like inflammation and cancer. However, due to physical constraints induced by the fiber bundle used for signal acquisition, obtaining pCLE images of high resolution is challenging. Consequently, in this study, we aim to improve pCLE image quality through the implementation of advanced post-processing techniques.

Here we propose an unsupervised single image super-resolution framework, which is free of using high-resolution pCLE images as reference and improves image quality significantly. The framework consists of a degradation module, a style transformation module and a super resolution module. In the degradation module, we propose an innovative distribution assumption module to randomize the fiber optic position distribution, enabling us to simulate the imaging principles of pCLE and create synthetic pCLE images for training.

With the integration of modules, both quantitative and qualitative analyses highlight the remarkable efficiency of our pipeline in super-resolving images compared to state-of-the-art methods. Our framework also demonstrates strong generalization capability, effectively mitigating the impact of pCLE system's intrinsic characteristics on image super-resolution. This feature is particularly advantageous as it allows the framework to circumvent redundant training when applied to various devices.

With the outstanding super-resolution and generalization capability, our proposed methodology enables clearer observation of image details and more accurate localization of micro structures, which contributes to precise identification of lesion areas and diagnostic accuracy enhancement.