Abstract
SIGNIFICANCE: Confocal endomicroscopic image stitching can expand the field of view and improve examination efficiency. However, due to interference from the rolling shutter effect, traditional stitching methods may produce misalignments, leading to structural distortion and artifacts. Suppressing the rolling shutter effect in confocal endomicroscopic images can effectively enhance stitching quality. AIM: We propose a Dual-Path Gaussian U-Net (DGU-Net)-based framework for confocal endomicroscopic image stitching. The parallel dual-encoder paths of DGU-Net extract Gaussian features and conventional features at different resolutions, respectively, achieving more precise gland segmentation masks. Based on these masks, we filter stable frames and optimize feature matching to effectively suppress rolling shutter interference and improve stitching quality. APPROACH: We annotated a segmentation dataset comprising 80 rat confocal laser endomicroscopy (CLE) images to train the segmentation network and validated the frame selection method's effectiveness in suppressing the rolling shutter effect on consecutively acquired rat CLE video sequences. The stitching results generated from the filtered stable image sequences were compared with conventional methods. RESULTS: Experimental results demonstrate that DGU-Net achieves superior performance with a Dice score of 85.17 on CLE datasets, significantly outperforming existing segmentation networks. Compared with Auto-Stitching, our method improves regional consistency across the panoramic image by eliminating artifacts caused by mismatches while delivering enhanced stitching accuracy and image quality. CONCLUSIONS: The proposed method effectively accomplishes confocal image stitching tasks, significantly enhancing endomicroscopic examination efficiency and contributing to improved diagnostic outcomes.