Super-resolution multimode fiber imaging with an untrained neural network.

Area of Science:

• Biomedical Optics
• Computational Imaging
• Machine Learning Applications

Background:

• Multimode fiber endoscopes enable miniaturized deep tissue imaging but typically have low spatial resolution and long acquisition times.
• Existing super-resolution techniques often rely on computationally intensive algorithms or machine learning requiring extensive training datasets and pre-calibration.

Purpose of the Study:

• To develop a fast and effective super-resolution imaging method for multimode fiber endoscopes.
• To overcome the limitations of traditional computational and machine learning approaches by eliminating the need for pre-training.

Main Methods:

• Implementation of an unsupervised learning framework utilizing untrained neural networks for image reconstruction.
• Addressing the ill-posed inverse problem inherent in multimode fiber imaging without prior data training.
• Theoretical and experimental validation of the proposed imaging methodology.

Main Results:

• Demonstrated significant enhancement in imaging quality compared to conventional methods.
• Achieved sub-diffraction spatial resolution, surpassing the diffraction limit of the optical system.
• Eliminated the need for large training datasets and lengthy pre-calibration procedures.

Conclusions:

• Untrained neural networks offer a practical and efficient solution for high-resolution imaging through multimode fibers.
• The proposed unsupervised learning approach advances minimally invasive deep tissue imaging capabilities.