EXPLORING DEEP LEARNING METHODS FOR LOW NUMERICAL APERTURE TO HIGH NUMERICAL APERTURE RESOLUTION ENHANCEMENT IN CONFOCAL MICROSCOPY

Author

Shashwat Patra

Date

2023

Document Type

Thesis

Degree Name

Master of Science

Department

Electrical & Computer Engineering

Committee Chair

Ana Doblas

Committee Member

Madhusudhanan Balasubramanian

Committee Member

Omar Skalli

Abstract

Confocal microscopy is a widely used tool that provides valuable morphological and functional information within cells and tissues. A major advantage of confocal microscopy is its ability to record multi-color and optically sectioned images. A major drawback to confocal microscopy is its diffraction-limited spatial resolution. Though techniques have been developed that break this limit in confocal microscopy, they require additional hardware or accurate estimates of the system’s impulse response (e.g., point spread function). Here we investigate two deep learning-based models, the cGAN and cycleGAN, trained with low-resolution (LR) and high-resolution (HR) confocal images to improve spatial resolution in confocal microscopy. Our findings conclude that the cGAN can accurately produce HR images if the training set contains images with a high signal-to-noise ratio. We have also found that the cycleGAN model has the potential to perform as the cGAN model but without the requirement of using paired inputs.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to ProQuest

Notes

Open Access

Recommended Citation

Patra, Shashwat, "EXPLORING DEEP LEARNING METHODS FOR LOW NUMERICAL APERTURE TO HIGH NUMERICAL APERTURE RESOLUTION ENHANCEMENT IN CONFOCAL MICROSCOPY" (2023). Electronic Theses and Dissertations. 3092.
https://digitalcommons.memphis.edu/etd/3092