PIIS0002939422003774
PIIS0002939422003774
PIIS0002939422003774
TIMOTHY P.H. LIN, HERBERT Y.H. HUI, ANNIE LING, POEMEN P. CHAN, RUYUE SHEN, MANDY O.M. WONG,
NOEL C.Y. CHAN, DEXTER Y.L. LEUNG, DEJIANG XU, MONG LI LEE, WYNNE HSU, TIEN YIN WONG,
CLEMENT C. THAM, AND CAROL Y. CHEUNG
• PURPOSE: To determine the relationship between base- ratio per SD decrease [95% confidence interval], 1.36
line retinal-vessel calibers computed by a deep-learning [1.01-1.82]) and CRVE (1.35 [1.01-1.80]) were associ-
system and the risk of normal tension glaucoma (NTG) ated with progressive RNFL thinning and narrower base-
progression. line CRAE (1.98 [1.17-3.35]) was associated with VF
• DESIGN: Prospective cohort study. deterioration.
• METHODS: Three hundred and ninety eyes from 197 • CONCLUSION: In this study, each SD decrease in the
patients with NTG were followed up for at least 24 baseline CRAE or CRVE was associated with a more
months. Retinal-vessel calibers (central retinal arteriolar than 30% increase in the risk of progressive RNFL thin-
equivalent [CRAE] and central retinal venular equiva- ning and a more than 90% increase in the risk of VF
lent [CRVE]) were computed from fundus photographs deterioration during the follow-up period. Baseline at-
at baseline using a previously validated deep-learning sys- tenuation of retinal vasculature in NTG eyes was as-
tem. Retinal nerve fiber layer (RNFL) thickness and vi- sociated with subsequent glaucoma progression. High-
sual field (VF) were evaluated semiannually. The Cox throughput deep-learning–based retinal vasculature anal-
proportional-hazards model was used to evaluate the re- ysis demonstrated its clinical utility for NTG risk as-
lationship of baseline retinal-vessel calibers to the risk of sessment. (Am J Ophthalmol 2023;247: 111–120. ©
glaucoma progression. 2022 Elsevier Inc. All rights reserved.)
• RESULTS: Over a mean follow-up period of 34.36 ±
5.88 months, 69 NTG eyes (17.69%) developed progres-
N
sive RNFL thinning and 22 eyes (5.64%) developed VF ormal tension glaucoma (NTG) is a
deterioration. In the multivariable Cox regression analy- common subgroup of primary open-angle glau-
sis adjusting for age, gender, intraocular pressure, mean coma in which the development and progression
ocular perfusion pressure, systolic blood pressure, ax- of the disease occur in the absence of elevated intraoc-
ial length, standard automated perimetry mean deviation, ular pressure (IOP).1 , 2 Hence, clinical risk assessment
and RNFL thickness, narrower baseline CRAE (hazard and management are particularly challenging because
the current glaucoma management is centered on IOP
reduction.3 , 4 NTG patients with higher risk of progression
Supplemental Material available at AJO.com.
may be identified by risk factors calculation.5 Nonetheless,
Clement C. Tham and Carol Y. Cheung contributed equally as senior au- the strategy of identifying this group of patients for close
thors and co-correspondence authors. monitoring and aggressive treatment remains theoretical
Accepted for publication September 23, 2022. and controversial. Furthermore, the pathogenesis of NTG
From the Department of Ophthalmology and Visual Sciences, the Chi-
nese University of Hong Kong (T.P.H.L., H.Y.H.H., A.L., P.P.C., R.S., remains unclear. Currently, evidence suggests a role of
M.O.M.W., N.C.Y.C., D.Y.L.L., C.C.T., C.Y.C.), Hong Kong, China; the “vascular theory” of glaucoma, which postulates that
Hong Kong Eye Hospital, Hong Kong, China; Lam Kin Chung, Jet King- circulatory insufficiency and ocular ischemia culminate in
Shing Ho Glaucoma Treatment and Research Centre, the Chinese Uni-
versity of Hong Kong(P.P.C., C.C.T., C.Y.C.), Hong Kong, China; De- retinal ganglion cell degeneration and glaucomatous visual
partment of Ophthalmology and Visual Sciences, Prince of Wales Hospi- field (VF) loss, in the pathogenic processes of NTG.6-8 In
tal (N.C.Y.C.), Hong Kong, China; Department of Ophthalmology, Hong support of this theory were observations that NTG eyes
Kong Sanatorium and Hospital (D.Y.L.L.), Hong Kong, China; School of
Computing, National University of Singapore (D.X., M.L.L., W.H.), Sin- had a higher prevalence of disc hemorrhage compared with
gapore; Singapore Eye Research Institute, Singapore National Eye Cen- other glaucoma subtypes, and patients with NTG were
tre (T.Y.W.), Singapore; Ophthalmology and Visual Sciences Academic more likely to suffer from systemic vascular diseases1 , 3 , 9 , 10
Clinical Programme, Duke-NUS Medical School (T.Y.W.), Singapore; Ts-
inghua Medicine, Tsinghua University (T.Y.W.), Beijing, China and microvascular dysregulation.11 , 12
Inquiries to Clement C. Tham and Carol Y. Cheung, Department of One potential method for the assessment of the vascu-
Ophthalmology and Visual Sciences, the Chinese University of Hong lar risk of an eye is to examine the state of the retinal
Kong, Kowloon, Hong Kong, China.; e-mail: clemtham@cuhk.edu.hk,
carolcheung@cuhk.edu.hk vasculature. We have previously demonstrated in a cross-
TABLE 1. Baseline Demographics of Eyes With and Without Normal Tension Glaucoma Progression
Variables Progressive RNFL Thinning Without Progressive RNFL Thinning P Value VF Deterioration Without VF Deterioration P Value
AMERICAN JOURNAL OF OPHTHALMOLOGY
CCT = central corneal thickness, CRAE = central retinal arteriolar equivalent, CRVE = central retinal venular equivalent, DBP = diastolic blood pressure, IOP = intraocular pressure, MD = mean
deviation, MOPP = mean ocular perfusion pressure, RNFL = retinal nerve fiber layer, SAP = standard automated perimetry, SBP = systolic blood pressure, VFI = visual field index.
MARCH 2023
FIGURE. Automated estimation of central retinal arteriolar caliber (CRAE) and central retinal venular caliber (CRVE) using
a deep-learning system (SIVA-DLS). The original retinal photograph captured by a 45° digital retinal camera (A). The color of
retinal arterioles is generally lighter, and the width of retinal arterioles is generally smaller than that of retinal venules. SIVA-DLS
automatically generates heat maps based on the features of the retinal vessels for prediction and estimation of CRAE (B) and CRVE
(C).
glaucoma progression (all Ps > .05). In comparison with the Table 3 shows the C-statistic for Cox regression models pre-
group without progressive RNFL thinning, the group with dicting glaucoma progression, comparing the model based
progressive RNFL thinning had lower MOPP (52.98 mm on baseline retinal-vessel calibers, age, gender, IOP, MOPP,
Hg vs 54.89 mm Hg, P = .049) and smaller CRAE (141.38 and SBP alone with the model based on all previously re-
μm vs 147.51 μm) and CRVE (226.23 μm vs 237.79 μm) ported risk factors. Models with inclusion of CRAE (C-
(P = .026 and P = .036, respectively) at baseline. In com- statistic, 0.702 vs 0.701, P = .595), CRVE (C-statistic,
parison with the group without VF deterioration, the group 0.700 vs 0.701, P = .726), or both CRAE and CRVE (C-
with VF deterioration had lower MOPP (49.54 mm Hg vs statistic, 0.703 vs 0.701, P = .242) showed comparable and
54.85 mm Hg, P = .001), lower SBP (124.00 mm Hg vs noninferior predictive discrimination for progressive RNFL
133.93 mm Hg) and DBP (72.95 mm Hg vs 79.55 mm thinning, compared with the model based on all previ-
Hg) (P = .001, P = .017 and P < .001, respectively), ously reported risk factors. Similar results were observed in
and smaller CRAE (138.73 μm vs 146.89 μm) and CRVE the model with inclusion of both CRAE and CRVE (C-
(228.39 μm vs 236.19 μm) at baseline, although the latter statistic, 0.85 vs 0.819, P = .026) for VF deterioration,
did not reach a statistically significant level (both Ps > .05). compared with the model based on all previously reported
Supplemental Figure 2 demonstrates baseline retinal-vessel risk factors. As shown in Supplemental Table 2, we further
calibers computed by SIVA-DLS and baseline demograph- compared the models with inclusion of baseline retinal-
ics between 2 eyes with and without glaucoma progression, vessel calibers and all previously reported risk factors with
respectively. the model based on all the previously reported risk factors
The relationships between baseline retinal-vessel cal- alone. The results were in line with those in Table 3, in
ibers and NTG progression are summarized in Table 2. In which models with inclusion of baseline retinal-vessel cal-
the multivariable Cox regression analysis, narrower base- ibers demonstrate statistically superior predictive discrimi-
line CRAE (HR [95% CI], 2.95 [1.20-7.23], first quartile nation for NTG progression to the model based on previ-
vs fourth quartile; HR per SD decrease, 1.36 [1.01-1.82]) ously reported risk factors (Supplemental Table 2).
and narrower baseline CRVE (HR per SD decrease, 1.35 Supplemental Figure 3 shows the prediction error curves
[1.01-1.80]) were independently associated with progressive of Cox regression models predicting glaucoma progression
RNFL thinning, after adjusting for age, gender, IOP, MOPP, determined by the presence of progressive RNFL thinning
SBP, AL, SAP MD, and RNFL thickness. Similarly, nar- and VF deterioration. The integrated Brier scores of these
rower baseline CRAE (HR, 5.57 [1.13-27.49], first quartile models at 48-month follow-up were calculated. Cox re-
vs fourth quartile; HR, 7.57 [1.56-36.73], second quartile gression models for NTG progression with inclusion of
vs fourth quartile; HR, 3.93 [1.03-14.98], third quartile vs previously reported risk factors and retinal-vessel calibers
fourth quartile; HR per SD decrease, 1.98 [1.17-3.35]) was (CRAE, CRVE, or both CRAE and CRVE) all had lower
associated with VF deterioration, but there were no statis- Brier scores (progressive RNFL thinning: 0.100; VF deterio-
tically significant associations between baseline CRVE and ration: 0.029), indicating better calibration compared with
VF deterioration. the model that included previously reported risk factors
The clinical predictive value of baseline retinal-vessel only (progressive RNFL thinning: 0.102; VF deterioration:
calibers for glaucoma progression was further evaluated. 0.030) and the model that included only baseline retinal-
Model 1 Model 2
CI = confidence interval, HR = hazard ratio, RNFL = retinal nerve fiber layer, VF = visual field.
Model 1 was unadjusted. Model 2 was adjusted for age, gender, intraocular pressure, mean ocular
perfusion pressure, systolic blood pressure, central corneal thickness, axial length, mean deviation,
and average RNFL thickness. Unless otherwise indicated, P values were determined on the basis
of caliber measures entered as a categorical variable (quartiles) and compared with the reference
quartile.
a
P values were determined on the basis of caliber measures entered as a continuous variable.
vessel calibers, age, gender, IOP, MOPP, and SBP (progres- findings revealed that baseline assessment of retinal-vessel
sive RNFL thinning: 0.103; VF deterioration: 0.031). calibers provides significant prognostic information for the
evaluation of the risk of NTG progression. In addition, the
results provided evidence to support the clinical application
of a DLS for high-throughput fundus photography analysis
DISCUSSION in NTG risk assessment.
In this study, we found that each SD decrease in the
In the current cohort study, baseline retinal-vessel calibers baseline CRAE or CRVE was associated with a more than
computed by a DLS were significantly associated with the 30% increase in the risk of progressive RNFL thinning
risk of glaucoma progression in NTG eyes, independent and a more than 90% increase in the risk of VF de-
of previously reported risk factors, such as age,32 , 33 gen- terioration during the follow-up period. To the best of
der,34 MOPP,36 SBP,37 AL,38 SAP MD,33 and RNFL thick- our knowledge, this is the first longitudinal study evalu-
ness39 at baseline. Furthermore, models based on baseline ating the association between baseline retinal-vessel cal-
retinal-vessel calibers demonstrated comparable or superior iber measurements computed by a DLS and the subse-
predictive discrimination of glaucoma progression to the quent risk of glaucoma progression in a cohort of patients
model based on previously reported risk factors only. These with NTG. Previously, evidence pinpointed an association
C-Statistic
Progressive RNFL thinning CRAE + Age + Gender + IOP + MOPP + SBP 0.702 0.701 +0.001 (+0.143) .595
Progressive RNFL thinning CRVE + Age + Gender + IOP + MOPP + SBP 0.700 0.701 −0.001 (−0.143) .726
Progressive RNFL thinning CRAE + CRVE + Age + Gender + IOP + MOPP + SBP 0.703 0.701 +0.002 (+0.285) .242
VF deterioration CRAE + Age + Gender + IOP + MOPP + SBP 0.848 0.819 +0.029 (+3.54) .085
VF deterioration CRVE + Age + Gender + IOP + MOPP + SBP 0.829 0.819 +0.01 (+1.22) .827
VF deterioration CRAE + CRVE + Age + Gender + IOP + MOPP + SBP 0.85 0.819 +0.031 (+3.79) .026
AL = axial length, CRAE = central retinal arteriolar caliber, CRVE = central retinal venular caliber, IOP = intraocular pressure, MOPP = mean
ocular perfusion pressure, RNFL = retinal nerve fiber layer, SAP = standard automated perimetry, SBP = systolic blood pressure, VF = visual
field.
a
Model with previously reported risk factors consisted of baseline age, gender, IOP, MOPP, SBP, AL, SAP mean deviation (MD), and RNFL
thickness.
between altered ocular hemodynamics and the incidence density at baseline were associated with progressive RNFL
of glaucoma.13 , 43 , 44 In the population-based Blue Moun- thinning, both supporting a clinically significant role of vas-
tains Eye study, glaucoma eyes were reported to be at least cular insufficiency in glaucoma development and progres-
2 times more likely to have retinal arteriolar narrowing sion.
than healthy eyes,22 and similar associations between glau- In this study, we also evaluated the application of DL-
coma and worse geometric measurements of the retinal vas- based retinal-vessel analysis in the clinical risk assessment
culature including decreased retinal-vessel tortuosity and of NTG. Previously, the evaluation of the retinal vascula-
branching angle were also demonstrated in the Singapore ture in glaucoma remains to be of limited utility in clini-
Malay Eye study.18 The attenuations of retinal-vessel pa- cal settings as evidence regarding the role of vascular pa-
rameters were further shown to be associated with worse rameters in risk assessment, prognosis, and management of
structural and functional glaucomatous measurements.13 , 45 glaucoma is lacking. Furthermore, computer programs de-
In these population-based studies, it was observed that glau- veloped for the analysis of fundus photographs or OCTA
coma were associated with both narrowing in retinal arte- images were highly labor intensive, time consuming, and
rioles and venules,20 , 22 , 46 and our current findings of the could be susceptible to human errors as substantial hu-
associations between baseline attenuation in CRAE and man input was involved.18-23 The implications of our re-
CRVE and glaucoma progression in NTG were largely cor- sults were 2-fold. First, we demonstrated that baseline reti-
roborative with earlier observations in the literature. These nal vasculature measurements obtained from SIVA-DLS,
findings on alterations in retinal vasculature in glaucoma when combined with basic demographical data (age, gen-
from fundus photography were in line with earlier ocular der, IOP, MOPP, and SBP), yielded statistically compara-
perfusion studies on Doppler imaging and fluorescein an- ble and noninferior outcomes for the prediction of NTG
giography14-17 and recent studies on retinal capillary net- progression, compared with the prediction model based on
works using optical coherence tomography angiography all previously reported risk factors obtained from exten-
(OCTA),13 , 43 , 44 , 47 which all underlined the relationship sive ophthalmic investigations (Table 3). Moreover, when
between retinal circulatory insufficiency and glaucoma. such baseline retinal vasculature measurements were added
In the past, studies on retinal vasculature and glaucoma into the prediction model based on all previously reported
were of largely cross-sectional design, and hence unable to risk factors, it further improved the predictive accuracy of
ascertain the temporal and causal relationships between de- NTG progression (Supplemental Table 2). This provided
generation in retinal vasculature and glaucoma develop- evidence to support the use of fully automated and highly ef-
ment. Findings from our longitudinal study provide novel ficient DL-based image analysis technology to compute reti-
and additional evidence to the elucidation of the vascular nal vascular parameters for glaucoma risk assessment from
theory6-8 in glaucoma development by demonstrating that fundus photographs. These findings are of particular clin-
degeneration in retinal vasculature heralds the advent of ical interest as fundus photography is a readily available
glaucoma progression in NTG. Our current findings also ophthalmic assessment, and by virtue of DL-based image
corroborate with the longitudinal results from OCTA re- analysis, this universal clinical assessment can yield signif-
ported in a recent study by Moghimi and associates,48 which icant prognostic information about the risk of progression
revealed that reduced optic nerve head and macular vessel in NTG to navigate ophthalmologists in identification of
Funding/Support: This work was supported by Health and Medical Research Fund, Hong Kong (Ref. No. 05162836 to C.C.T.), and General Research
Fund, Hong Kong (Ref. No. 14107516 to C.C.T.). The funding organization had no role in the design or conduct of this research.
Financial Disclosures: The authors declare no conflict of interest. All authors attest that they meet the current ICMJE criteria for authorship.