ConNeCT: weakly supervised corneal confocal microscopy image inpainting network based on a diffusion model.

Abstract

Quantitative analysis of the corneal nerve morphology using corneal confocal microscopy (CCM) has shown significant potential for diagnosing a range of neurodegenerative diseases. However, images acquired using current CCM devices are often affected by various artifacts, which can compromise the accuracy of parameter measurements. In this study, we proposed ConNeCT, i.e., a weakly supervised image inpainting network designed specifically for CCM images. ConNeCT took a raw artifact-laden image along with a coarse user-provided mask as input and performed end-to-end image restoration. The framework comprised three main components: (1) a lightweight guided diffusion model based on a denoising diffusion probabilistic model (DDPM) enhanced with deformable convolutions for improved feature extraction, (2) a U-Net-based auxiliary segmentation model, and (3) an improved DDPM resampling algorithm. The resampling process iteratively leveraged information from artifact-free regions to reconstruct structurally consistent images guided by gradient signals from the segmentation model to better preserve nerve fiber structures. An evaluation on a manually annotated dataset demonstrated that the proposed method outperformed existing approaches (RePaint, MCG, DDNM, and DeqIR), achieving state-of-the-art results with SSIM = 0.9838, PSNR = 17.68, HD = 13.74, MSD = 6.30, and MAE = 14.80. To the best of our knowledge, our study outcome is the first deep learning-based method specifically developed for CCM image inpainting.