Faculty Bibliography

PURPOSE:The presence of imbalanced datasets in medical applications can negatively affect deep learning methods. This study aims to investigate how the performance of convolutional neural networks (CNNs) for glaucoma diagnosis can be improved by addressing imbalanced learning issues through utilizing glaucoma suspect samples, which are often excluded from studies because they are a mixture of healthy and preperimetric glaucomatous eyes, in a semi-supervised learning approach. METHODS:A baseline 3D CNN was developed and trained on a real-world glaucoma dataset, which is naturally imbalanced (like many other real-world medical datasets). Then, three methods, including reweighting samples, data resampling to form balanced batches, and semi-supervised learning on glaucoma suspect data were applied to practically assess their impacts on the performances of the trained methods. RESULTS:The proposed method achieved a mean accuracy of 95.24%, an F1 score of 97.42%, and an area under the curve of receiver operating characteristic (AUC ROC) of 95.64%, whereas the corresponding results for the traditional supervised training using weighted cross-entropy loss were 92.88%, 96.12%, and 92.72%, respectively. The obtained results show statistically significant improvements in all metrics. CONCLUSIONS:Exploiting glaucoma suspect eyes in a semi-supervised learning method coupled with resampling can improve glaucoma diagnosis performance by mitigating imbalanced learning issues. TRANSLATIONAL RELEVANCE:Clinical imbalanced datasets may negatively affect medical applications of deep learning. Utilizing data with uncertain diagnosis, such as glaucoma suspects, through a combination of semi-supervised learning and class-imbalanced learning strategies can partially address the problems of having limited data and learning on imbalanced datasets.

PURPOSE/UNASSIGNED:The structural changes measured by optical coherence tomography (OCT) are related to functional changes in visual fields (VFs). This study aims to accurately assess the structure-function relationship and overcome the challenges brought by the minimal measurable level (floor effect) of segmentation-dependent OCT measurements commonly used in prior studies. METHODS/UNASSIGNED:We developed a deep learning model to estimate the functional performance directly from three-dimensional (3D) OCT volumes and compared it to the model trained with segmentation-dependent two-dimensional (2D) OCT thickness maps. Moreover, we proposed a gradient loss to utilize the spatial information of VFs. RESULTS/UNASSIGNED:Our 3D model was significantly better than the 2D model both globally and pointwise regarding both mean absolute error (MAE = 3.11 + 3.54 vs. 3.47 ± 3.75 dB, P < 0.001) and Pearson's correlation coefficient (0.80 vs. 0.75, P < 0.001). On a subset of test data with floor effects, the 3D model showed less influence from floor effects than the 2D model (MAE = 5.24 ± 3.99 vs. 6.34 ± 4.58 dB, P < 0.001, and correlation 0.83 vs. 0.74, P < 0.001). The gradient loss improved the estimation error for low-sensitivity values. Furthermore, our 3D model outperformed all prior studies. CONCLUSIONS/UNASSIGNED:By providing a better quantitative model to encapsulate the structure-function relationship more accurately, our method may help deriving VF test surrogates. TRANSLATIONAL RELEVANCE/UNASSIGNED:DL-based VF surrogates not only benefit patients by reducing the testing time of VFs but also allow clinicians to make clinical judgments without the inherent limitations of VFs.

PURPOSE/UNASSIGNED:Lamina cribrosa (LC) deformation is hypothesized to play a major role in glaucoma pathogenesis. The purpose of this study was to determine in vivo how varying intraocular pressure (IOP) under fixed intracranial pressure (ICP), and vice versa, deforms the pore paths throughout the LC volume. METHODS/UNASSIGNED:Spectral-domain optical coherence tomography scans of the optic nerve head were acquired from healthy adult rhesus monkeys under different pressures. IOP and ICP were controlled with gravity-based perfusion systems into the anterior chamber and lateral ventricle, respectively. IOP and ICP were modulated from baseline to high (19-30 mmHg) and highest (35-50 mmHg) levels while maintaining a fixed ICP of 8 to 12 mmHg and IOP of 15 mmHg, respectively. After three-dimensional registration and segmentation, the paths of pores visible in all settings were tracked based on their geometric centroids. Pore path tortuosity was defined as the measured distance divided by the minimal distance between the most anterior and posterior centroids. RESULTS/UNASSIGNED:The median pore tortuosity at baseline varied among the eyes (range, 1.16-1.68). For the IOP effect under fixed ICP (six eyes, five animals), two eyes showed statistically significant increased tortuosity and one showed a decrease (P < 0.05, mixed-effects model). No significant change was detected in three eyes. When modulating ICP under fixed IOP (five eyes, four animals), a similar response pattern was detected. CONCLUSIONS/UNASSIGNED:Baseline pore tortuosity and the response to acute pressure increase vary substantially across eyes. TRANSLATIONAL RELEVANCE/UNASSIGNED:LC pore path tortuosity could be associated with glaucoma susceptibility.

Objective: Although artificial intelligence (AI) models may offer innovative and powerful ways to use the wealth of data generated by diagnostic tools, there are important challenges related to their development and validation. Most notable is the lack of a perfect reference standard for glaucomatous optic neuropathy (GON). Because AI models are trained to predict presence of glaucoma or its progression, they generally rely on a reference standard that is used to train the model and assess its validity. If an improper reference standard is used, the model may be trained to detect or predict something that has little or no clinical value. This article summarizes the issues and discussions related to the definition of GON in AI applications as presented by the Glaucoma Workgroup from the Collaborative Community for Ophthalmic Imaging (CCOI) US Food and Drug Administration Virtual Workshop, on September 3 and 4, 2020, and on January 28, 2022. Design: Review and conference proceedings. Subjects: No human or animal subjects or data therefrom were used in the production of this article. Methods: A summary of the Workshop was produced with input and approval from all participants. Main Outcome Measures: Consensus position of the CCOI Workgroup on the challenges in defining GON and possible solutions. Results: The Workshop reviewed existing challenges that arise from the use of subjective definitions of GON and highlighted the need for a more objective approach to characterize GON that could facilitate replication and comparability of AI studies and allow for better clinical validation of proposed AI tools. Different tests and combination of parameters for defining a reference standard for GON have been proposed. Different reference standards may need to be considered depending on the scenario in which the AI models are going to be applied, such as community-based or opportunistic screening versus detection or monitoring of glaucoma in tertiary care. Conclusions: The development and validation of new AI-based diagnostic tests should be based on rigorous methodology with clear determination of how the reference standards for glaucomatous damage are constructed and the settings where the tests are going to be applied.

BACKGROUND:Spectral-domain (SD-) optical coherence tomography (OCT) can reliably measure axonal (peripapillary retinal nerve fiber layer [pRNFL]) and neuronal (macular ganglion cell + inner plexiform layer [GCIPL]) thinning in the retina. Measurements from 2 commonly used SD-OCT devices are often pooled together in multiple sclerosis (MS) studies and clinical trials despite software and segmentation algorithm differences; however, individual pRNFL and GCIPL thickness measurements are not interchangeable between devices. In some circumstances, such as in the absence of a consistent OCT segmentation algorithm across platforms, a conversion equation to transform measurements between devices may be useful to facilitate pooling of data. The availability of normative data for SD-OCT measurements is limited by the lack of a large representative world-wide sample across various ages and ethnicities. Larger international studies that evaluate the effects of age, sex, and race/ethnicity on SD-OCT measurements in healthy control participants are needed to provide normative values that reflect these demographic subgroups to provide comparisons to MS retinal degeneration. METHODS:Participants were part of an 11-site collaboration within the International Multiple Sclerosis Visual System (IMSVISUAL) consortium. SD-OCT was performed by a trained technician for healthy control subjects using Spectralis or Cirrus SD-OCT devices. Peripapillary pRNFL and GCIPL thicknesses were measured on one or both devices. Automated segmentation protocols, in conjunction with manual inspection and correction of lines delineating retinal layers, were used. A conversion equation was developed using structural equation modeling, accounting for clustering, with healthy control data from one site where participants were scanned on both devices on the same day. Normative values were evaluated, with the entire cohort, for pRNFL and GCIPL thicknesses for each decade of age, by sex, and across racial groups using generalized estimating equation (GEE) models, accounting for clustering and adjusting for within-patient, intereye correlations. Change-point analyses were performed to determine at what age pRNFL and GCIPL thicknesses exhibit accelerated rates of decline. RESULTS:The healthy control cohort (n = 546) was 54% male and had a wide distribution of ages, ranging from 18 to 87 years, with a mean (SD) age of 39.3 (14.6) years. Based on 346 control participants at a single site, the conversion equation for pRNFL was Cirrus = -5.0 + (1.0 × Spectralis global value). Based on 228 controls, the equation for GCIPL was Cirrus = -4.5 + (0.9 × Spectralis global value). Standard error was 0.02 for both equations. After the age of 40 years, there was a decline of -2.4 μm per decade in pRNFL thickness ( P < 0.001, GEE models adjusting for sex, race, and country) and -1.4 μm per decade in GCIPL thickness ( P < 0.001). There was a small difference in pRNFL thickness based on sex, with female participants having slightly higher thickness (2.6 μm, P = 0.003). There was no association between GCIPL thickness and sex. Likewise, there was no association between race/ethnicity and pRNFL or GCIPL thicknesses. CONCLUSIONS:A conversion factor may be required when using data that are derived between different SD-OCT platforms in clinical trials and observational studies; this is particularly true for smaller cross-sectional studies or when a consistent segmentation algorithm is not available. The above conversion equations can be used when pooling data from Spectralis and Cirrus SD-OCT devices for pRNFL and GCIPL thicknesses. A faster decline in retinal thickness may occur after the age of 40 years, even in the absence of significant differences across racial groups.

PURPOSE/UNASSIGNED:Lamina cribrosa (LC) deformations caused by elevated intraocular pressure (IOP) are believed to contribute to glaucomatous neuropathy and have therefore been extensively studied, in many conditions, from in vivo to ex vivo. We compare acute IOP-induced global and local LC deformations immediately before (premortem) and after (postmortem) sacrifice by exsanguination. METHODS/UNASSIGNED:The optic nerve heads of three healthy monkeys 12 to 15 years old were imaged with spectral-domain optical coherence tomography under controlled IOP premortem and postmortem. Volume scans were acquired at baseline IOP (8-10 mm Hg) and at 15, 30, and 40 mm Hg IOP. A digital volume correlation technique was used to determine the IOP-induced three-dimensional LC deformations (strains) in regions visible premortem and postmortem. RESULTS/UNASSIGNED:Both conditions exhibited similar nonlinear relationships between IOP increases and LC deformations. Median effective and shear strains were, on average, over all eyes and pressures, smaller postmortem than premortem, by 14% and 11%, respectively (P's < 0.001). Locally, however, the differences in LC deformation between conditions were variable. Some regions were subjected premortem to triple the strains observed postmortem, and others suffered smaller deformations premortem than postmortem. CONCLUSIONS/UNASSIGNED:Increasing IOP acutely caused nonlinear LC deformations with an overall smaller effect postmortem than premortem. Locally, deformations premortem and postmortem were sometimes substantially different. We suggest that the differences may be due to weakened mechanical support from the unpressurized central retinal vessels postmortem. TRANSLATIONAL RELEVANCE/UNASSIGNED:Additional to the important premortem information, comparison with postmortem provides a unique context essential to understand the translational relevance of all postmortem biomechanics literature.

BACKGROUND AND OBJECTIVES/OBJECTIVE:Recent studies have suggested that inter-eye differences (IEDs) in peripapillary retinal nerve fiber layer (pRNFL) or ganglion cell+inner plexiform (GCIPL) thickness by spectral-domain optical coherence tomography (SD-OCT) may identify people with a history of unilateral optic neuritis (ON). However, this requires further validation. Machine learning classification may be useful for validating thresholds for OCT IEDs and for examining added utility for visual function tests, such as low-contrast letter acuity (LCLA), in the diagnosis of people with multiple sclerosis (PwMS) and for unilateral ON history. METHODS:Participants were from 11 sites within the International Multiple Sclerosis Visual System (IMSVISUAL) consortium. pRNFL and GCIPL thicknesses were measured using SD-OCT. A composite score combining OCT and visual measures was compared individual measurements to determine the best model to distinguish PwMS from controls. These methods were also used to distinguish those with history of ON among PwMS. ROC curve analysis was performed on a training dataset (2/3 of cohort), then applied to a testing dataset (1/3 of cohort). Support vector machine (SVM) analysis was used to assess whether machine learning models improved diagnostic capability of OCT. RESULTS:Among 1,568 PwMS and 552 controls, variable selection models identified GCIPL IED, average GCIPL thickness (both eyes), and binocular 2.5% LCLA as most important for classifying PwMS vs. controls. This composite score performed best, with AUC=0.89 (95% CI 0.85, 0.93), sensitivity=81% and specificity=80%. The composite score ROC curve performed better than any of the individual measures from the model (p<0.0001). GCIPL IED remained the best single discriminator of unilateral ON history among PwMS (AUC=0.77, 95% CI 0.71,0.83, sensitivity=68%, specificity=77%). SVM analysis performed comparably to standard logistic regression models. CONCLUSIONS:A composite score combining visual structure and function improved the capacity of SD-OCT to distinguish PwMS from controls. GCIPL IED best distinguished those with history of unilateral ON. SVM performed as well as standard statistical models for these classifications. CLASSIFICATION OF EVIDENCE/METHODS:The study provides Class III evidence that SD-OCT accurately distinguishes multiple sclerosis from normal controls as compared to clinical criteria.

Purpose : Multiple sublayers of retina can be visualized with visible light (vis-) OCT.However, image quality can be compromised due to patient movement, cataracts, small pupil size, and light scattering causing haziness and variability in signal to noise ratio in individual A-scans and in entire B-scans.The purpose of this study was to examine the effect of conventional and deep neural network dehazing techniques on the visibility and quantitative assessment of retinal sub-layers on vis-OCT images. Methods : 9 healthy and 5 glaucoma subjects were scanned 3 times during one session.Scanning was done on the superior nasal side of para-foveal region,1.5 mm from the fovea with a 3D speckle reduction raster scanning protocol(3x3x1.6 mm with 8192x16x1024 samplings) using a prototype vis-OCT system.16 A-scan lines were averaged to reduce speckle noise.Gray-scale image dehazing guided by depth information and pretrained Dehazenet deep model following deep convolutional neural network with residual learning(DnCNN) were applied on original B-scans.Quality improvement were evaluated using quality index(QI) and contrast to noise ratio(CNR) on dehazed B-scans.For each subject, the dehazed B-scan of Dehazenet and DnCNN from a fixed location adjacent to the fovea were selected.The distances between each of 3 bright inner plexiform layers(IPL) and retinal pigment epithelium(RPE) sublayers were segmented manually for thickness measurements using a 8 A-scan averaged profile(Fig.).Coefficient of variations (CVs) were calculated to assess the measurement repeatability of the sublayers on original and dehazed B-scans. Results : Healthy and glaucoma subjects were age 45.67+/-11.7and 59.60+/-13.4(p=0.07,t-test),visual field mean deviation(MD)-1.55 to1.20 dB,and from -26.42 to -7.70dB(p= 0.003,Wilcoxon),global mean circumpapillary retinal nerve fiber layer(RNFL)thickness 96.33+/-12.20 and 59.80+/-9.09mm(p<0.001,Wilcoxon),respectively.Dehazed B-scans obtained by deep models have statistically significant better QI and CNR(Table1).Overall intra-subject CVs showed significantly improved reproducibility on all measured sub-layers of dehazed B-scans compared to original scans for all subjects(Tables 2,3). Conclusions : Vis-OCT image quality can be improved using deep neural network dehazing model resulting in higher reproducible thickness measurements of retinal sublayers within subjects in dehazed B-scans

Purpose : Deep learning of optical coherence tomography (OCT) may help discriminate glaucomatous eyes from healthy controls. However, the underlying decision making processes remain unclear. Recently, through computing class activation maps, our feature agnostic artificial intelligence of OCT images using a 3D convolutional neural network identified the optic nerve head (ONH) and its surrounding regions as structures significantly associated with glaucoma classification (PMID: 31260494). To pursue their contributions further, here we analyzed the optic nerve morphology from OCT and magnetic resonance imaging (MRI) in a subset of glaucoma and healthy subjects. Methods : Nine early glaucoma, 12 advanced glaucoma, and 4 healthy control subjects underwent spectral-domain OCT at 30x30x2 mum³ and 3-Tesla anatomical MRI at 1x1x1mm³ . Maximum intensity projection was applied to en-face OCT scans at the ONH (Fig. 1). The areas of the ONH [inner regions of interest, (ROI)] and surrounding regions (outer ROIs) visible in OCT were measured using global thresholding in ImageJ. One-way ANOVAs with post-hoc Tukey's tests were performed on the inner and outer ROIs between the 3 groups. Also, a Pearson correlation analysis was performed between the ROI areas in OCT and optic nerve volume extracted from MRI between the eye and optic chiasm. Results : For OCT of the ONH, significant group effect was observed for the areas in the inner ROIs (ANOVA: F= 7.823, p=0.00133). Post-hoc analyses revealed a significant difference between healthy controls and advanced glaucoma (p=0.0082) and between early and advanced glaucoma (p=0.0057) but no significance between healthy controls and early glaucoma (p=0.80) (Fig. 2A). No significant group effect was observed in the outer ROIs (ANOVA: F=0.004, p=0.996) (Fig. 2B). There was a negative correlation between the inner ROI area in OCT and optic nerve volume in MRI (R=-0.47, p=0.0011) (Fig. 2C). Conclusions : The ONH tissues visible on OCT appeared to contribute more than their surrounding regions to distinguishing between glaucomatous eyes and healthy eyes. The negative correlation between ONH area in OCT and optic nerve volume in MRI suggested the need to further understand the interactions between ONH and deeper brain structures in glaucoma. (Figure Presented)

Purpose : The lamina cribrosa (LC) is hypothesized to be the site of initial axonal damage in glaucoma with the peripapillary retinal nerve fiber layer (RNFL) thickness is widely used as a standard metric for quantifying this damage. The purpose of this study was to determine in vivo changes in the microstructure of the LC in eyes of non-human primates (NHP) with naturally occurring RNFL thinning. Methods : Spectral-domain OCT scans (Leica, Chicago, IL) of the optic nerve head (ONH) were acquired in vivo from a colony of 50 adult rhesus monkeys, suspected of having high prevalence of naturally occurring glaucoma. The circumpapillary global and quadrant RNFL thickness was analyzed using a custom automated segmentation software. From the set of 100 eyes, the 10 eyes with the thinnest global RNFL values were selected as the study group, while 10 eyes with RNFL values around the 50 percentile were used as the control group. A previously described automated segmentation algorithm was used for LC microstructure analysis. The LC microstructure was analyzed globally and in the th following volumetric sectors: quadrants, central and peripheral lamina, and 3 depth slabs (anterior, middle, posterior; Figure). Beam thickness/pore diameter ratio (BPR) and connective tissue volume fraction (CTVF: beam volume/total volume) were calculated globally and in sectors. Results : 20 eyes (15 animals) were analyzed (Table 1). While no significant difference was detected between groups for age, weight or disc size, the study group had significantly thinner RNFL than the control group (p<0.01). The study group had significantly larger BPR and CTVF compared with the control group (Table 2). Significant sectoral differences between study and control group RNFL thickness were noted for BPR and CTVF in the nasal and temporal quadrants, central LC, and in LC depth. Across eyes, the global RNFL thickness was moderately negatively correlated only with the global CTVF (lower RNFL thickness associated with higher CTVF; r² =0.63, p=0.045). Conclusions : Eyes with thinner circumpapillary RNFL had thicker LC BPR and CTVF globally and in various sectors when compared to eyes with normal RNFL thickness. Whether these LC changes are the cause of RNFL damage or the result of remodeling of the LC requires further investigation. (Figure Presented)