Rolling shutter-resistant confocal endomicroscopy image stitching via dual-path Gaussian U-Net

Area of Science:

• Medical imaging
• Computational pathology
• Biomedical engineering

Background:

• Confocal endomicroscopic image stitching expands the field of view but is hindered by rolling shutter effects, causing misalignments and artifacts.
• Effective suppression of rolling shutter effects is crucial for improving stitching quality in confocal endomicroscopy.

Purpose of the Study:

• To propose a novel framework, Dual-Path Gaussian U-Net (DGU-Net), for enhancing confocal endomicroscopic image stitching.
• To effectively suppress rolling shutter interference and improve the accuracy and quality of stitched images.

Main Methods:

• A DGU-Net framework with parallel dual-encoder paths was developed to extract Gaussian and conventional features at different resolutions for precise gland segmentation.
• Stable frames were filtered, and feature matching was optimized based on segmentation masks to mitigate rolling shutter effects.
• A segmentation dataset of 80 rat confocal laser endomicroscopy (CLE) images was annotated for training, and the frame selection method was validated on CLE video sequences.

Main Results:

• DGU-Net achieved a Dice score of 85.17 on CLE datasets, outperforming existing segmentation networks.
• The proposed method significantly improved regional consistency and eliminated artifacts compared to conventional Auto-Stitching.
• Enhanced stitching accuracy and image quality were demonstrated in experimental results.

Conclusions:

• The DGU-Net based method effectively addresses challenges in confocal image stitching, particularly the rolling shutter effect.
• This approach significantly enhances endomicroscopic examination efficiency.
• The improved stitching quality contributes to better diagnostic outcomes in clinical applications.