«2015 Computational methods to model disease and genetic effects on optic nerve head structure Mark Allen Christopher University of Iowa Copyright ...»
University of Iowa
Iowa Research Online
Theses and Dissertations
Computational methods to model disease and
genetic effects on optic nerve head structure
Mark Allen Christopher
University of Iowa
Copyright 2015 Mark Allen Christopher
This dissertation is available at Iowa Research Online: http://ir.uiowa.edu/etd/1959
Christopher, Mark Allen. "Computational methods to model disease and genetic effects on optic nerve head structure." PhD (Doctor
of Philosophy) thesis, University of Iowa, 2015.
Follow this and additional works at: http://ir.uiowa.edu/etd Part of the Biomedical Engineering and Bioengineering Commons
COMPUTATIONAL METHODS TO MODEL DISEASE AND GENETIC EFFECTS
ON OPTIC NERVE HEAD STRUCTUREby Mark Allen Christopher A thesis submitted in partial fulfillment of the requirements for the Doctor of Philosophy degree in Biomedical Engineering in the Graduate College of The University of Iowa December 2015 Thesis Supervisors: Professor Michael D. Abràmoff Associate Professor Todd E. Scheetz Copyright by
MARK ALLEN CHRISTOPHER2015 All Rights Reserved Graduate College The University of Iowa Iowa City, Iowa
CERTIFICATE OF APPROVAL_______________________
PH.D. THESIS _______________
This is to certify that the Ph.D. thesis of Mark Allen Christopher has been approved by the Examining Committee for the thesis requirement for the Doctor of Philosophy degree in Biomedical Engineering at the December 2015 graduation.
Thesis Committee: ___________________________________
Michael D. Abràmoff, Thesis Supervisor ___________________________________
Todd E. Scheetz, Thesis Supervisor ___________________________________
John H. Fingert ___________________________________
Mona K. Garvin ___________________________________
Joseph M. Reinhardt
ACKNOWLEDGMENTSI would first like to thank my advisors Dr. Todd Scheetz and Dr. Michael Abràmoff for their mentorship throughout my time as an undergraduate and graduate student. I’d also like to thank everyone involved in the Scheetz Lab, WIVR, and CBCB for providing the environment I’ve needed to develop as a scientist. Finally, I’d like to thank Bhavna Antony. Her help in revising this thesis, discussions and suggestions for all of my projects, and the support she’s provided over the years have been invaluable.
Glaucoma is a leading cause of blindness throughout the world and is estimated to affect 80 million by 2020. This disease causes progressive loss of vision and, left untreated, can lead to complete blindness. With treatment, however, disease progression can be slowed dramatically. This makes early detection and intervention crucial in preserving the vision of affected individuals.
Onset and progression of glaucoma are associated with structural changes to an anatomical feature known as the optic nerve head (ONH). The ONH is the site of attachment between the retina and the optic nerve that carries all visual information to the brain. As glaucoma progresses, characteristic changes related to cell death and loss of vision can be observed in the three-dimensional structure of the ONH. A common modality used to observe these changes is stereo fundus imaging. This modality captures three-dimensional information via stereo imaging and is commonly used in clinical settings to diagnose and monitor glaucoma. A limitation of using stereo fundus images is the need for review by glaucoma specialists to identify disease related features of ONH structure. Further, even when expert evaluation is possible, the subjective nature of the process can lead due large discrepancies in the evaluations and resultant clinical decisions. The work presented here seeks address these concerns by providing automated, computational tools that can be used to characterize ONH structure.
Specifically, this thesis outlines the development of computational methods for inferring three-dimensional information from stereo fundus images and identifying objective, quantitative measurements of ONH structure. The resulting computational tools were applied to image and clinical data collected from a large cohort of individuals to identify hidden relationships between ONH structure, clinical measurements, and glaucoma. These tools were then applied to develop methods for estimating the impact of individual genetic factors on the ONH. Finally, using a longitudinal dataset collected over
Early detection is a crucial aspect of care in the treatment of glaucoma. This progressive disease causes irreversible loss of vision and can lead to complete blindness.
However, with early intervention, disease progression can be dramatically slowed and vision can be retained. This work presents data-driven methods to identify structural changes associated with glaucoma and aid in early detection of the disease.
The focus of the methods presented here is to analyze the three-dimensional shape of an anatomical structure known as the optic nerve head (ONH). The ONH is the attachment site of the optic nerve to the retina with a characteristic shape that often undergoes changes during the development and progression of glaucoma. Observation of the ONH is a standard part of clinical assessments for glaucoma. By applying statistical and computational techniques to a large dataset of medical images and clinical measurements, biologically and clinically important features of ONH structure were identified.
Specifically, methods for quantifying ONH structure based on medical images were developed and the resulting measurements were found to significantly increase accuracy in predicting development of glaucoma. Further methods that incorporated genetic information were developed and used to identify significant relationships between ONH shape and genetics. Finally, longitudinal data captured over several years was analyzed to identify time-dependent ONH changes associated with disease.
LIST OF TABLES
LIST OF FIGURES
Primary Open Angle Glaucoma
Optic Nerve Head Structure
Statistical Shape Modeling of Retinal Structure
Genetic Associations with POAG
3. FUNDUS IMAGE PROCESSING
Optic Nerve Head Localization
4. ANALYSIS OF BASELINE ONH STRUCTURE
Structural Endophenotype Identification
Structural Endophenotype Evaluation
Clinical Measurement Associations
Comparison to Expert ONH Review
5. DISCOVERY OF GENETIC ASSOCIATIONS WITH BASELINE ONH STRUCTURE
Genetic Contributions of Known Glaucoma Polymorphisms
Known Glaucoma Polymorphism Identification
Structural Endophentoype Identification
Genetic Feature Estimation
Genetic Feature Evaluation
Genome-Wide Structural Endophenotype Associations
Preprocessing and Quality Control
6. LONGITUDINAL ANALYSIS OF ONH STRUCTURE
Longitudinal Fundus Image Registration
Baseline Annotation and Image Preprocessing
Image Registration and Evaluation
Individual Depth Map Analysis
Depth Inference and Structural Endophenotype Identification..............86 Clinical Measurement Prediction
Longitudinal Changes to Structural Endophenotypes
Longitudinal Measurement Preprocessing
Longitudinal Features in Disease Detection
Longitudinal Features in Disease Prediction
4.1 Summary of baseline data and POAG outcomes for the OHTS cohort
4.2 Significant associations between OHTS baseline measurements and STEPs..........47
4.3 Summary of expert grading and comparisons to STEPs.
5.1 Set of SNPs previously associated with POAG.
5.2 Significant associations between STEPs and SNPs from across the genome..........71
6.1 OHTS cohort demographics, longitudinal measurements, and disease status..........75
6.2 The registration parameters and preprocessing steps that were evaluated...............80
6.3 Significant associations between longitudinal measurements and STEPs...............87
Figure 2.1 (A) Cross-section of the human eye with the location of the ONH highlighted (inset).
Illustration courtesy of National Eye Institute (https://nei.nih.gov/photo/). (B) The ONH as it appears in fundus images with the nasal (N), temporal (T), superior (S), and inferior (I) directions labeled.
(C) A three-dimensional rendering of the ONH illustrating typical structure............8
2.2 Example of a healthy ONH and some common abnormalities. (A) A normal ONH region with the cup (blue circle) and disc (red circle) outlined. The ratio of the diameters of these circles corresponds to the CDR. (B) An ONH region exhibiting localized thinning (notching) in the superior quadrant (blue arrow).
(C) An example of peripapillary atrophy altering the pigment around the ONH (red arrow).
3.1 The segmented window overlaid onto an example fundus image. This window was used to compute features at each pixel location. Features were computed based on each quadrant (I – IV) individually as well as the entire window.
3.2 The set of PCA-derived filters that were used to augment optic disc predictive features. Each of these filters was applied at three different scales to compute features used to estimate ONH location.
3.3 An example input image is shown along with images illustrating the selected features used in the automated ONH localization method. The features are shown in the order that they were chosen in during greedy forward feature selection.
3.4 The effect varying the number of nearest neighbors, k, on model performance. The lowest observed error was achieved by setting k equal to 11.
3.5 The ONH localization model applied to example fundus images. The input images (left) are shown along model output (right). Final point estimations (blue) and human annotated truth (red) is also shown.
3.6 In stereo photography, the disparity (the difference between the position of a point in each image) is inversely proportional to the points three-dimensional depth. The scene points P1 and P2 differ only in depth (d1 d2). Dashed lines indicate the projection of each point onto the images I1 and I2. Their disparities are given by x1 and x2. Note that as depth decreases from d1 to d2, the corresponding disparity increases from x1 to x2.
3.7 The depth inference algorithm uses a multi-scale image representation. In this representation, the scale dimension extends from coarse (down-sampled, Gaussian-blurred) to fine (full resolution, un-blurred) images.
3.9 Example depth maps inferred from stereo (top) and renderings of the corresponding ONH structure (bottom).
3.10 Comparisons of ONH region depth inferred from stereo fundus images to depth measured via OCT.
4.1 Example baseline stereo fundus images from the OHTS dataset. The format differences between different pairs are a result of the different camera types that were used. Despite quality assurance several possible sources of error can be observed with the images. These include over/under illumination, the presence of bright glare artifacts, and differences in focus within an image pair.
4.2 Illustration of the fundus and depth processing procedure. (A) Input stereo images. (B) The extracted ONH region stereo pair. (C) The raw depth map, the edge-cropped map, and the smoothed map. (D) Renderings of the ONH structure corresponding to the depth map.
4.3 Gray-scale representation of the ten PCA-based STEPs used to model ONH structure shown with the percent of variance in depth data explained by each.
Collectively, these features accounted for 95% of the variance observed................46
4.4 STEP features estimating the contribution of demographic and clinical variables to ONH structure.
4.5 Area under receiver operating characteristic curves for incident POAG prediction using combinations of demographic (age, sex, ethnicity), clinical (HCDR, VCDR, IOP, CCT, PSD, refraction), and STEP features
4.6 Variations to ONH structure capture by STEPs. (A) Gray-scale representations of the first five STEPs. (B) Illustration of the change to ONH structure induced by increasing the contribution of a single STEP. The median structure (left) is shown along with the median altered to exhibit an extreme value of the STEP. (C) The same figure shown as heat maps to indicate depth.
5.1 STEP features identified using PCA applied to the genotyped subset of the OHTS cohort (n = 1054).
5.2 Gray-scale representations of the estimated effect of 19 POAG-related SNPs on ONH structure. Each is shown along with the gene in which the SNPs occur. The * indicates the relationship between the estimated effect and SNP genotype is significant.
5.3 Summary of the associations between genetic ONH structural features and POAG. (A) The full list of identified genetic features associated with POAG.
(B) The effect on ONH structure for some of the most significantly associated features. The gray-scale feature is along with the effect on structure that results from increasing the influence of the feature.