Zero-Shot Super-Resolution With a Physically-Motivated Downsampling Kernel for Endomicroscopy.

Area of Science:

• Medical Imaging
• Computational Imaging
• Biomedical Engineering

Background:

• Super-resolution (SR) methods using convolutional neural networks (CNNs) have advanced medical imaging.
• Endomicroscopy imaging quality is often limited by the lack of ground truth high-resolution (HR) images for training and assessment.
• Unsupervised SR methods are being explored to overcome these limitations.

Purpose of the Study:

• To develop a novel zero-shot super-resolution (ZSSR) approach for endomicroscopy that does not require ground truth HR images.
• To improve non-reference image quality in endomicroscopy using self-supervised learning.
• To tailor SR methods to the specific characteristics of endomicroscopy data.

Main Methods:

• A novel ZSSR pipeline was designed for self-supervised, non-reference image quality improvement.
• A physically-motivated Voronoi downscaling kernel was introduced to account for endomicroscope's fibre-based sampling.
• Realistic noise patterns and video sequences were utilized for enhanced self-supervised learning.

Main Results:

• Ablation studies confirmed the contributions of the downscaling kernel and noise simulation.
• The ZSSR method demonstrated superior performance in image quality compared to baseline methods on synthetic and original data.
• User studies with experts and non-experts showed ZSSR reconstructions were competitive with supervised SR and preferred by experts.

Conclusions:

• The proposed ZSSR method effectively improves endomicroscopy image quality without requiring ground truth HR images.
• The tailored approach, incorporating a Voronoi kernel and realistic noise, addresses the unique challenges of endomicroscopy.
• ZSSR offers a promising alternative for image enhancement in settings where ground truth data is unavailable.