Faculty Bibliography

Purpose : To improve the performance of the feature agnostic AI-based glaucoma detection algorithm by evaluating an uncertainty score for each prediction. Methods : We previously developed a 5-layer 3D Convolutional Neural Network (CNN) in using the OCT scans from both eyes of 134 healthy, 779 glaucoma patients on a Cirrus HDOCT scanner (200x200 ONH Cubes; Zeiss, Dublin CA). In our analysis, we excluded scans with signal strength less than 7 and downsampled the volumes to 64x64x128 voxels. Uncertainty of AI models can be estimated by computing the effect of randomly ignoring a set of parameters within the network. We randomly zeroed 5% of each of the 5 convolutional layers and computed the entropy in the final score over 20 forward passes. The performance of the approach was assessed using a 10-fold cross validation study. Results : Over the 10-folds, the model showed an AUC of 0.91+/-0.027. In analysing the uncertainty and the probabilistic scores generated by the model (Softmax function) for one fold (see Fig. 1), we observed that a threshold of 0.8 can be used to flag 75% of the false positives and false negatives for further review. On the other hand, only 25% of the healthy controls and 20% of glaucoma patients showed an uncertainty score above that threshold. Fig. 2 summarises the overall uncertainties scores and indicates that low scores are associated with the correctly identified cases while the errors show higher uncertainty scores. Conclusions : The quantitative uncertainty measure provides supplementary information to clinicians and can be used to flag difficult cases automatically. Given that the dataset used in this work is highly imbalanced (more positive cases compared to normal cases) the uncertainty score for true negative cases is significantly higher compared to true positive cases. We expect to achieve lower uncertainty scores for normal cases if more data for normal eyes are available. The uncertainty analysis presented here may aid clinical interpretations of AI-based glaucoma detection outcomes. A separate study will be run to measure this improvement and compare the result with experts' level of uncertainty

Purpose : The uncertainty quantification of segmentation results is critical for understanding the reliability of the segmentation model. The purpose of this study is to investigate the effect of deep learning-based segmentation with uncertainty measurement in the relationship between RNFL thickness and visual field mean deviation (MD) Methods : Optical coherence tomography (OCT) scans were acquired from both eyes on 634 glaucoma patients, 404 glaucoma suspects, and 49 healthy controls using commercial OCT device (Cirrus HD-OCT, 200x200 Optic Disc Cubes; Zeiss, Dublin, CA). All subjects had visual field (VF) tests at each visit (Humphrey VF, SITA 24-2 test; Zeiss). A segmentation model was trained using Bayesian deep learning for voxel-wise segmentation of RNFL layer in OCT volume and compute the voxel-wise uncertainty of the segmentation output. The higher uncertainty denotes the unreliability of the segmentation and vice versa and it allows the determination of erroneous segmentation at test time. Uncertainty-guided global mean of the RNLF thickness (RNFL-Umean) was then computed by discarding the voxels with erroneous segmentation labels with higher uncertainty during the thickness computation. Also, the global mean of the RNLF thickness (RNFLmean;) was computed without taking uncertainty into account. Pearson correlation coefficient between RNFLU and MD was computed and compared with the Pearson correlation coefficient between RNFLmean;and MD. Results : The proposed RNFL-Umean;gave stronger correlation with MD than RNFLmean;The Pearson correlation coefficients were (0.67 (RNFL-Umean;) vs 0.63 (RNFL mean;; p<0.001) for glaucoma subjects, (0.56 vs 0.53 ;p=0.01) for glaucoma suspects and (0.08 vs 0.01; p=0.21) for normal subjects. Conclusions : The proposed uncertainty-guided computation of RNFL thickness showed improved correlation with the visual field MD. This demonstrates that segmentation uncertainty can be used to reduce the effect of inaccurate segmentation in computing the RNFL thickness. This also shows that uncertainty-guided computation of RNFL thickness is a better predictor of visual function than the normal RNFL thickness computed without using uncertainty

Purpose : Alterations in retinal oxygen metabolism are implicated in blindness causing diseases, such as diabetic retinopathy and glaucoma. Therefore, a non-invasive clinical tool to assess oxygen saturation (sO2) in retinal vessels is desirable. Recent development of visible-light optical coherence tomography (vis-OCT) enabled non-invasive sO measurements in retinal blood vessels at micrometer-scale resolution by threedimensional (3D) spectroscopic analysis. Nevertheless, such measurements are susceptible to spectral contaminants from the complex retina anatomy and vis-OCT signal detection and processing, decreasing measurement reliability. To overcome limitations posed by spectral contaminants, we developed adaptive-spectroscopic OCT (AS-OCT), a processing technique that enables non-invasive, 3D, environment-independent sO measurements in the human retina. Methods : We used vis-OCT to image the retinas of 18 healthy volunteers. Light exposure in the eye was < 250 muW and imaging acquisition time was 5 sec. We used adaptive spectroscopic OCT (AS-OCT) to identify and remove contaminants from retinal tissues, chromatic aberrations, and spectrally-dependent roll-off. Then, we automatically selected the optimal depths in the vessel for sO measurement. Finally, we measured the attenuation spectrum in the blood vessel and used a least-squares regression fit with known spectra to determine the sO value. Results : We measured sO in 125 unique retinal vessels near the optic disc (vessel diameters ranging from 37 mum to 176 mum). Major arteries had sO2= 97.9 +/-2.9 % (mean +/-standard deviation) (n = 36), small arteries (diameter < 100 had sO2= 93.2 +/-5.0 % (n = 36), and veins had sO2= 58.5 +/-4.3 % (n = 53). Repeated measurement standard deviations were 2.21% and 2.32% for all arteries and veins, respectively. Fig. 1 shows an oximetry map of the optic disc in the retina of a healthy 23 year-old volunteer. Conclusions : AS-OCT enables environment-independent retinal oximetry in the clinical setting. Repeatability in arteries and veins < 2.5 % indicates robust measurements that are promising for clinical use

Purpose : Current Optical Coherence Tomography (OCT) denoising techniques mainly focus on denoising 2-dimensional (2D) B-scans, especially for deep learning (DL) methods, which assume spatial integrity among neighboring samplings. However, OCT signal is essentially one dimensional (1D), and eye movements during scanning often violate the assumption. The purpose of this study was to see if 1D denoising is a feasible approach for clinical OCT imaging. Methods : 3D SD-OCT data within 6x6x2mm volumes centered on optic nerve head were obtained from 121 eyes (79 patients). Clean reference scans were constructed by registering and averaging 6 OCT scans obtained on the same day using ANTs software. As shown in Figure 1, we used a 5-layer U-shape net (UNet) for both 2D denoiser (Figure 1.(a)) and 1D denoiser (Figure1.(b)). In addition, both 2D and 1D approaches are combined by using the 2D denoised B-scan as a mask to selectively remove high signal peaks in the 1D denoised B-scan (Figure 1.(c)). Peak signal-to-noise ratio (PSNR) and contrast-to-noise ratio (CNR) were calculated to compare the performance. Results : Subjectively, the 2D denoiser generated smoother edges but tended to oversmooth textual information within the retinal layers, while the 1D denoiser preserved more textual information but caused more jittering at retinal edges due to the lack of structural information from neighboring A-scans. Quantitatively, the 1D denoiser showed similar PSNR to the 2D denoiser, while outperforming in CNR (PSNR: 31.20 (1D) V.S. 31.20 dB (2D), p=0.963; CNR: 4.23 (1D) V.S. 3.90 dB (2D), p<0.001, paired t-test, Table 1). The combined denoiser further improved CNR (4.39 (combined) V.S. 3.90 dB (2D), p<0.001). Combining 1D and 2D denoisers, the denoised B-scan showed more continuous edges compared to the 1D denoiser and did not lose the texture information compared to the 2D denoiser (Figure 2). Conclusions : Quantitatively, 1D denoiser performance is as good as 2D denoiser or even better. A combination of 1D and 2D approaches may provide well-balanced image enhancement in clinical applications

Purpose : Mutations in optineurin (OPTN) are associated with familial normal tension glaucoma and other neurodegenerative diseases. It remains unclear how OPTN loss or mutation alters visual function during aging. Here, we used transgenic mouse models and in vivo assessments to test the hypothesis that OPTN dysfunction contributes to progressive visual impairment through a toxic gain of function mechanism. Methods : Mice with C57BL/6 background were used (Fig 1): wildtype (WT; n=19), homozygous OPTN knock-out (mOPTN-KO; n=13), hemizygous mouse E50K OPTN knock-in (mE50K-het; n=8), homozygous mouse E50K OPTN knock-in (mE50K homoz; n=10), and human E50K OPTN bacterial artificial chromosome overexpression (hE50K BAC; n=6) (PMID: 31076632, 25818176). Intraocular pressure (IOP), total retinal thickness (TRT), visual acuity (VA), and contrast sensitivity (CS) were measured at 6, 12, and 18 months of age in the same mice using the TonoLab rebound tonometer, Bioptigen spectral-domain optical te uses cookies. By continuing to use our website, you are agreeing to coherence tomography imaging, and OptoMotry optokinetic virtual reality system respectively. Left and right eye data were averaged and analyzed using ANOVAs followed by posthoc tests between genotype and age groups, as well as linear regressions for VA versus contrast threshold (CT). Results : Our longitudinal study of the same mice during the aging process showed that IOP remained normal between 10-15 mmHg (Fig 2A). Small to no difference in TRT over time or compared to WT was observed (Fig 2B). mE50K-homoz, mE50K-het, and hE50K BAC mice exhibited greater age-dependent decline in VA and CT than WT or mOPTN-KO mice (Fig 2C, 2D, 2E). In contrast, mOPTN-KO mice showed preservation of VA and CT over time compared to WT. Consistently, mice with one copy of E50K OPTN (mE50K het) experienced less deterioration of VA and CT compared to mice with two copies (mE50K homoz) or mild overexpression (hE50K BAC). Conclusions : Depsite limited IOP and TRT changes between age and genotype groups, E50K OPTN was associated with differential age-dependent visual impairment (greater for CS than VA). Surprisingly, OPTN deficiency preserved visual function such that CS in knockout mice was better than WT mice. Our results suggest visual loss associated with E50K OPTN is due to a toxic gain of function mechanism, and that suppression of OPTN might constitute a therapeutic strategy for glaucomatous neurodegeneration

Purpose : Lamina cribrosa (LC) deformation is hypothesized to be a major cause of glaucoma. The LC undergoes different forms of stress both anteriorly from intraocular pressure (IOP), as well as posteriorly and circumferentially from subarachnoid cerebrospinal fluid pressure (CSFP) and the sclera. The purpose of this study was to determine possible in vivo changes in the path of the lamina pores under different IOP settings while maintaining fixed CSFP. Methods : Spectral-domain OCT scans (Leica, Chicago, IL) of the optic nerve head (ONH) were acquired in vivo under different pressure settings from healthy rhesus monkeys. IOP was controlled using a gravity-based perfusion system through a needle inserted into the anterior chamber. CSFP was maintained at the baseline opening pressure via gravitybased perfusion system through cannulation of the brain's lateral ventricle (range 8- 12mmHg). Scans were acquired at baseline IOP (15mmHg), high (30 mmHg) and very high IOP (40-50 mmHg) and registered in 3D. Pores from shared regions were automatically segmented using a previously described segmentation algorithm. The path of each pore was tracked based on the calculated geometric centroid of each pore. The tortuosity of each pore path was defined as the total actual distance of the centroid path divided by the minimal distance between the first (most anterior) and last (most posterior) pore centroids. Results : 7 eyes from 6 healthy adult Rhesus macaque were analyzed. The mean value of the pore path tortuosity varies between eyes at baseline IOP levels (range: 1.16-1.68; Table). Two main overall patterns of pore path tortuosity were detected in response to increased IOP at fixed CSFP: 4 eyes became more tortuous (M2, M5, M8, M11); in the rest of the eyes (M6 OD, M6 OS, M10) the pore paths remained either unchanged or showed a variable response. No statistically significant change (p > 0.05) was observed in this small sample in either the subject-specific analysis or the analysis of the pooled combined values of the pore path tortuosity. Conclusions : Baseline pore tortuosity as well as the response of the pores to acute IOP increase varies between eyes. Further investigation is warranted to determine if these differences are associated with glaucoma susceptibility

Purpose : Dampened cerebrovascular reactivity (CVR) impairs blood delivery to brain regions. Multiple studies have suggested a role of the basal forebrain (BF) in glaucoma (PMID: 31242454; ARVO 2020: 4336). However, CVR changes in BF with glaucoma severity have yet to be explored. Recently, relative CVR (rCVR) mapping was introduced using resting-state functional MRI (rsfMRI) without gas challenges. Here, we investigate rCVR changes in the visual cortex and basal forebrain with glaucoma severity. Methods : Normal (n=22), early-stage (n=18), and advanced-stage (n=19) glaucoma patients underwent anatomical MRI and rsfMRI. The optic nerve and optic chiasm volumes were measured using ImageJ. rCVR maps and regional rCVR values were generated and extracted with MriCloud online tool. Age, optical coherence tomography measurements, and Humphrey visual field perimetry were obtained from clinical records. Results are presented as mean+/-SEM. One-way ANOVA followed by post-hoc Bonferroni's multiple comparisons test, and trend analysis were applied. Results : Demographics, clinical ophthalmic assessments, and volumetric MRI assessments illustrated the characteristics of the anterior visual pathways in the normal control, earlystage glaucoma and advance-stage glaucoma groups (Fig. 1). The averaged rCVR map from normal controls is consistent with previous studies (PMID: 27888058, 32278094) (Fig. 2). Advanced-stage glaucoma patients [1.03+/-0.03 relative unit (r.u.); p<0.05] have significantly lower rCVR in the visual cortex compared to normal controls (1.20+/-0.06 r.u.; early-stage: 1.09+/-0.05 r.u.), and exhibit a decreasing trend in more severe disease. These corroborate with a previous Doppler ultrasound study (PMID: 23816432). Interestingly, rCVR in BF has an increasing trend with severity. No significant rCVR change was observed in the somatosensory cortex. Conclusions : Visual cortical rCVR decreases with glaucoma severity, while rCVR in the basal forebrain increases with severity. Our results verify visual cortical CVR reduction in glaucoma patients and further solidify that the basal forebrain plays a role in glaucoma

Purpose : Compare the speed of glaucoma progression as measured by global visual field (VF) and optical coherence tomography (OCT) metrics. Methods : Glaucoma suspect, pre-perimetric glaucoma (PPG) and perimetric glaucoma (PG) participants of Advanced Imaging for Glaucoma study, who had at least 7 visits with visual field (VF) and OCT ONH scan, were analyzed. Severity was staged by the modified Hodapp-Parrish-Anderson Criteria. Eyes with significant cataract progression were excluded. Spectral-domain OCT (RTVue, Optovue) and VF testing were performed every 6 months. The nerve fiber layer (NFL) thickness was measured from ONH scan. The NFL mean deviation (MD) were VF-equivalent dB-scale quantities based on sectorwise nonlinear regression of NFL thickness with VF deviation using cross-sectional data over a wide range of glaucoma severity. Linear regression was used to estimate the glaucoma progression speed. Results : Seventy-five glaucoma suspect eyes (VF MD-0.1+/-1.2dB), 160 PPG eyes (0.2+/-1.3 dB), 77 early PG eyes (-0.9+/-1.6dB) and 20 moderate+severe PG (-10.8+/-3.2dB) eyes were analyzed. The follow-up duration was 54 months +/- 8 months (mean +/- SD). For both VF MD and NFL MD, the speed of progression increased monotonically with glaucoma severity (Table 1). For overall NFL thickness, the progression speed was greatest in the PPG and early PG stages, but slowed down at the moderate+severe stages. The ratios of progression speed for NFL thickness relative to VF were significantly different between stages (p<0.006, one-way ANOVA). The ratios of progression speed for NFL MD relative to VF MD was generally slower (0.58-0.72), but not significantly different across disease stages (p=0.08). The progression speed of both NFL-MD and VF-MD were associated with baseline parameters (faster progression in eyes with more severe disease at baseline), while progression speed of NFL thickness was not (Table 2). Conclusions : Compared to VF MD, NFL thickness tends to overestimate the progression speed in the early stages of glaucoma and underestimate it in the later stages. Clinicians should be aware of the discrepancy in the apparent speed of disease progression as measured by structural and functional metrics, which strongly depend on the stage of disease severity. Converting the NFL thickness profile to NFL MD may provide a progression metric more consistent with VF MD over a wider range of glaucoma severity

Purpose : To assess the repeatability of estimates of mean deviation (MD) and visual field index (VFI) obtained from an automated deep-learning approach that analysed raw OCT volumes. Methods : OCT scans were acquired from both eyes of 138 healthy, 743 glaucoma suspects and 941 glaucoma patients (Cirrus HD-OCT scanner, 200x200 ONH Cubes, Zeiss, Dublin CA). The scans were acquired at multiple visits, with two or more scans acquired at each visit. Scans with signal strength < 7 were discarded, giving us a total of 19,208 OCT scans. A subset of 5207 eyes (total of 10,414 scans) had repeat scans of that met the inclusion criteria. 24-2 Humphrey visual field (VF) tests were administered at each visit. A single convolutional neural network was trained to estimate the MD and VFI (dual outputs) from downsampled OCT volumes (50x50x128 voxels). The network consisted of 5 convolutional layers, followed by a global average pooling layer and dual outputs to enable the simultaneous estimation of MD and VFI. A mean squared error loss was used to train the network using an Adam optimiser over a total of 200 epochs. A 10-fold cross-validation scheme was used, where the dataset was divided into 10 non-overlapping folds (~182 subjects per fold) - trained on 8-folds, validated on one and tested on one. Each subject was limited to a unique fold. The performance of the method was assessed by computing the median error and interquartile range. The repeatability was assessed using a set of 5207 OCT scans that had repeats available. Results : The median absolute error (Q1, Q3) for the estimates of MD and VFI were 1.66 (0.79, 2.99) dB and 3.01 (1.48, 6.63) %, respectively. In the reproducibility test, the Pearson's correlation coefficient was 0.91 (CI: [0.91, 0.92]) and 0.91 (CI: [0.90, 0.92]), for MD and VFI, respectively. The median absolute difference between the repeated estimates for MD and VFI were 0.53 (0.21, 0.51) dB and 1.17 (0.45, 1.14)%, respectively. Conclusions : The deep-learning based approach for estimating visual field test parameters shows repeatability better than expected test-to-test variability

Purpose : Repetitive transorbital alternating current stimulation (rtACS) is an application of weak electric current near the eyes used in vision rehabilitation of optic neuropathies (ON). Conceptually rtACS entrains neuronal oscillations, augmenting neuronal function. In subjects with ON we evaluated whether rtACS influenced visual structure and function. Methods : 34 subjects with ON enrolled in a prospective trial underwent comprehensive ophthalmic evaluation, visual field (VF) 24-2 and 10-2 tests (Humphrey Field Analyzer) and OCT (Cirrus HD-OCT) retinal nerve fiber layer (RNFL) and ganglion cell inner plexiform layer (GCIPL) thicknesses at baseline and follow-up (FU) visits. Subjects received rtACS 30-to 45-minutes daily for 10 days. Sham subjects (n=4) underwent the same procedures but received no current. Point-by-point analyses of VF total deviation (TD) values were conducted between rtACS and sham groups. Regression analyses determined rate of change for each TD point per eye (significant points with positive rate of change defined as improved, negative rate of change as progressed; insignificant rate of change as no change) and the association between RNFL and GCIPL between groups. Results : The number of FU visits with VF tests ranged 2 to 7, with no significant differences detected between rtACS vs sham groups' FU duration. No significant differences were detected between groups' baseline VF 24-2 and 10-2 mean deviation (MD) values (Table 1). The average numbers of improved points (VF 10-2) and progressed points (VF 24-2) were greater for rtACS while the average number of no change points was greater for sham (VF 24-2, p0.05, Table 1). Further analysis of FU duration determined a significant interaction with rtACS; number of improved points (VF 10-2) and progressed points (VF 24-2, p<0.02) were not sustained over time. No significant differences were detected in average RNFL and GCIPL thicknesses between groups. Conclusions : Preliminary analyses of the effect of rtACS in ON indicate initial improvement but not a clear benefit over time. Detection of differences between rtACS vs sham groups may be biased due to the small sham sample and range of FU duration as VF test-to-test variability is known to increase with worsening VF MD. Future analyses will assess interim effect at early vs late FU time points to evaluate the role of rtACS in vision rehabilitation