Adversarial training with cycle consistency for unsupervised super-resolution in endomicroscopy.

Area of Science:

• Medical Imaging
• Computational Imaging
• Deep Learning

Background:

• Endomicroscopy is crucial for real-time diagnostics and interventional guidance, aiding in cancer detection.
• Acquisition constraints limit endomicroscopy image quality due to low pixel counts.
• Super-Resolution (SR) techniques can enhance image quality, but supervised methods require paired low-resolution (LR) and high-resolution (HR) images.

Purpose of the Study:

• To develop an unsupervised SR framework for endomicroscopy images.
• To address the challenge of limited paired LR/HR data in endomicroscopy.
• To improve the quality and diagnostic utility of endomicroscopy images.

Main Methods:

• An unsupervised SR framework utilizing adversarial deep neural networks.
• Implementation of a physically-inspired cycle consistency loss.
• Leveraging unpaired HR images from any domain to transfer quality to LR images.

Main Results:

• Quantitative analysis on 238 endomicroscopy video sequences from 143 patients.
• Demonstrated ability to produce convincing super-resolved images.
• Validation through a Mean Opinion Score (MOS) study confirming image quality improvements.

Conclusions:

• The proposed unsupervised SR framework effectively enhances endomicroscopy image quality without requiring paired data.
• This approach is valuable for improving real-time tissue characterization and cancer detection in clinical settings.
• The method offers a viable solution for enhancing medical images where paired data is scarce.