Jay Shenoy

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2022-159

May 20, 2022

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-159.pdf

The human retina contains a mosaic of light-sensitive photoreceptor cells that capture visual stimuli. Determining the structure of the retina is important for ophthalmology and vision science, as well as for emerging display technologies that operate at the cellular level. Adaptive optics scanning laser ophthalmoscopy (AOSLO) and optical coherence tomography (AO-OCT) are two techniques for imaging the retina at high resolution in 2D and 3D, respectively. Both techniques scan the eye over a set period of time, and as a result produce images that contain distortions arising from the motion of the eye during acquisition. Prior methods for AOSLO and AO-OCT image processing have been reference-based, generating maps of the retina by registering the acquired image frames against a motion-corrected reference image and averaging these frames together. These methods rely on heuristics for estimating eye motion, resulting in inaccurate motion traces that engender mapping artifacts. In this work, we introduce an optimization-based framework for retina tracking and mapping that directly solves for the most likely trace of eye motion that occurred during the recording session. Our framework, R-SLAM, uses inter-frame feature correspondences and a novel convex optimization algorithm to compute the optimal motion solution, where optimality is defined in a probabilistic sense. By directly solving the inverse problem of calculating the likeliest map and motion trace that gave rise to the retina recording, we produce retina maps of higher quality than those found in prior work. This report includes research on AOSLO image processing that was part of a recent publication as well as subsequent work on applying R-SLAM to the AO-OCT regime that was conducted during the master's program. R-SLAM's success in both the AOSLO and AO-OCT domains indicates its utility as a theoretical and practical foundation that opens up new avenues for research on optimization-based retinal image processing.

Advisors: Ren Ng

---

BibTeX citation:

@mastersthesis{Shenoy:EECS-2022-159,
    Author= {Shenoy, Jay},
    Editor= {Ng, Ren and Roorda, Austin},
    Title= {Optimization Methods for Tracking and Mapping the Human Retina},
    School= {EECS Department, University of California, Berkeley},
    Year= {2022},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-159.html},
    Number= {UCB/EECS-2022-159},
    Abstract= {The human retina contains a mosaic of light-sensitive photoreceptor cells that capture visual stimuli. Determining the structure of the retina is important for ophthalmology and vision science, as well as for emerging display technologies that operate at the cellular level. Adaptive optics scanning laser ophthalmoscopy (AOSLO) and optical coherence tomography (AO-OCT) are two techniques for imaging the retina at high resolution in 2D and 3D, respectively. Both techniques scan the eye over a set period of time, and as a result produce images that contain distortions arising from the motion of the eye during acquisition. Prior methods for AOSLO and AO-OCT image processing have been reference-based, generating maps of the retina by registering the acquired image frames against a motion-corrected reference image and averaging these frames together. These methods rely on heuristics for estimating eye motion, resulting in inaccurate motion traces that engender mapping artifacts. In this work, we introduce an optimization-based framework for retina tracking and mapping that directly solves for the most likely trace of eye motion that occurred during the recording session. Our framework, R-SLAM, uses inter-frame feature correspondences and a novel convex optimization algorithm to compute the optimal motion solution, where optimality is defined in a probabilistic sense. By directly solving the inverse problem of calculating the likeliest map and motion trace that gave rise to the retina recording, we produce retina maps of higher quality than those found in prior work. This report includes research on AOSLO image processing that was part of a recent publication as well as subsequent work on applying R-SLAM to the AO-OCT regime that was conducted during the master's program. R-SLAM's success in both the AOSLO and AO-OCT domains indicates its utility as a theoretical and practical foundation that opens up new avenues for research on optimization-based retinal image processing.},
}

EndNote citation:

%0 Thesis
%A Shenoy, Jay 
%E Ng, Ren 
%E Roorda, Austin 
%T Optimization Methods for Tracking and Mapping the Human Retina
%I EECS Department, University of California, Berkeley
%D 2022
%8 May 20
%@ UCB/EECS-2022-159
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-159.html
%F Shenoy:EECS-2022-159