Depixelation and image restoration with meta-learning in fiber-bundle-based endomicroscopy

Area of Science:

• Medical Imaging
• Computational Biology
• Machine Learning

Background:

• Fiber-bundle-based endomicroscopy is crucial for in vivo imaging.
• Honeycomb artifacts degrade image quality and hinder accurate diagnosis.
• Existing methods for artifact removal are often time-consuming or require specific prior information.

Purpose of the Study:

• To develop an efficient meta-learning algorithm for removing honeycomb artifacts.
• To restore high-contrast, noise-free images from fiber-bundle endomicroscopy.
• To create a versatile method applicable to various imaging conditions without prior tissue or fiber distribution knowledge.

Main Methods:

• A meta-learning algorithm utilizing two sub-networks for feature extraction was developed.
• Meta-training was performed on limited simulated data to ensure generalization.
• The algorithm was tested on USAF targets and living mouse tissue images.

Main Results:

• The algorithm successfully removed honeycomb artifacts and restored high-contrast images.
• Restored images exhibited reduced pixilated noise.
• The method achieved efficient artifact removal in a shorter time compared to existing techniques.
• The approach demonstrated applicability across diverse fiber-bundle endomicroscopy conditions.

Conclusions:

• The developed meta-learning algorithm offers an efficient solution for honeycomb artifact removal in fiber-bundle endomicroscopy.
• This method significantly improves image quality for biological tissue visualization.
• Its adaptability makes it a valuable tool for a wide range of endomicroscopy applications.