Thesis
Automated retinal layer segmentation in mouse OCT using deep learning and graph search
University of Iowa
Master of Science (MS), University of Iowa
Spring 2025
DOI: 10.25820/etd.007861
Abstract
Optical coherence tomography (OCT), a non-invasive imaging technique, produces cross-sectional information of the retina and quantitative measurements of the retinal structural changes, which provide valuable insights into disease progression. Traditionally, retinal layer thicknesses in mice are measured using manual caliper tracing on OCT displays. While manual tracing can be effective, it is time-consuming, prone to human errors, and often shows intra-observer and inter-observer variability. This thesis seeks to overcome these limitations by evaluating a two-step three-dimensional deep learning graph-search (DLGS) approach, that was originally designed to segment retinal layers in human macular OCT scans, even in cases of severe retinal ganglion cell layer complex (RGC) thinning, for mouse OCT.
Thus, the primary objective of this thesis was to adapt the DLGS approach to segment OCT scans from mouse models and assess the advantages of automation. To evaluate the robustness of the DLGS model, structural changes in RGC thickness in a mouse model of optic neuritis (ON) were quantified using the following: an inbuilt Bioptigen caliper, a publicly available mouse layer segmentation model with averaging masks, DLGS model with the same masks, and DLGS with those from manually corrected surfaces created by a human tracer using a 3D image viewer. Averaging masks were applied to regions where inbuilt caliper measured thickness located at four regions (nasal, superior, temporal, and inferior), 400 µm from the optic nerve head. Measured thicknesses from each model were then compared across all grid configurations by assessing signed and unsigned differences, as well as evaluating consistency using Pearson correlation coefficients.
The DLGS model and manual caliper measurements showed strong agreement across masks that averaged thickness over broader retinal regions, with an average Pearson correlation of 0.787. Moreover, the mean signed difference and mean absolute difference were −3.57 ± 3.08 µm and 4.00 ± 2.54 µm, respectively, indicating that the DLGS deviated from the manual caliper measurements by only 0 to 5 voxels across images that were 1024 voxels in height. Additionally, Pearson correlation coefficients between the DLGS model and DLGS with 3D surface correction exceeded 0.9 across all comparisons, suggesting strong agreement, with absolute differences ranging from 0.28 to 0.33 µm. Meanwhile, the publicly available model showed modest correlations with manual calipers (r = −0.471 to −0.145) and mean differences of ∼ 39 ± 7.5 µm. These results demonstrate the DLGS model’s robustness in layer segmentation, as evidenced by its low variability compared to the 3D-corrected DLGS. The caliper measurement results further highlight the potential impact of human errors and observational bias. Overall, this suggests that current methods for RGC complex segmentation offer a promising alternative to manual tracing. By improving these models, researchers testing experimental medicine using murine models can benefit from automated workflows, which will eliminate the need for manual retinal layer thickness measurements. Furthermore, segmented layer labels can be more easily obtained, facilitating the development of additional deep-learning models that can potentially predict future vision loss in human or non-human subjects based on en face images or thickness maps derived from segmented layers.
Details
- Title: Subtitle
- Automated retinal layer segmentation in mouse OCT using deep learning and graph search
- Creators
- Yun Jae Choi
- Contributors
- Mona K Garvin (Advisor)Terry A Braun (Committee Member)Michael J Schnieders (Committee Member)
- Resource Type
- Thesis
- Degree Awarded
- Master of Science (MS), University of Iowa
- Degree in
- Biomedical Engineering
- Date degree season
- Spring 2025
- DOI
- 10.25820/etd.007861
- Publisher
- University of Iowa
- Number of pages
- xi, 51 pages
- Copyright
- Copyright 2025 Yun Jae Choi
- Language
- English
- Date submitted
- 04/07/2025
- Description illustrations
- illustrations, tables, graphs
- Description bibliographic
- Includes bibliographical references (pages 46-51).
- Public Abstract (ETD)
- In this thesis, I developed an enhanced method for analyzing optical coherence tomography (OCT) scans of mouse retinas. OCT provides high resolution, cross sectional images that are essential for measuring the thickness of retinal layers, par- ticularly the retinal ganglion cell (RGC) complex, which is crucial for monitoring disease progression. Current measurement techniques such as using built in software, manual tracing, or laboratory-based methods are often time-consuming, inconsistent, and may fail to detect subtle changes across the retina. To overcome these limitations, I adapted a two-step, three-dimensional deep learning graph search (DLGS) algorithm for auto- matically analyzing the RGC layer in mice with optic neuritis, a condition that causes substantial RGC thinning and serves as a model for human retinal diseases. I evaluated my DLGS model using standard averaging masks against three mea- surement strategies: the Bioptigen caliper tool, a publicly available mouse segmenta- tion model, and an expert corrected version of my DLGS approach. Results demon- strated that my DLGS algorithm closely matched the Bioptigen measurements when the thickness was averaged across regions typically differing by only a few microm- eters, while the publicly available model showed considerably larger discrepancies. Notably, my DLGS measurements were nearly identical to the expert corrected anal- yses, confirming the accuracy of my method. This automated approach significantly reduces analysis time and human error, po- tentially accelerating retinal research and enabling the development of more sophisti- cated predictive models for disease progression based on precise layer segmentation.
- Academic Unit
- Roy J. Carver Department of Biomedical Engineering
- Record Identifier
- 9984830826102771
Metrics
1 File views/ downloads
12 Record Views