Glucose tolerance and insulin resistance/sensitivity associate with retinal layer characteristics: the LIFE-Adult-Study.

AIMS/HYPOTHESIS: As the prevalence of insulin resistance and glucose intolerance is increasing throughout the world, diabetes-induced eye diseases are a global health burden. We aim to identify distinct optical bands which are closely related to insulin and glucose metabolism, using non-invasive, high-resolution spectral domain optical coherence tomography (SD-OCT) in a large, population-based dataset. METHODS: The LIFE-Adult-Study randomly selected 10,000 participants from the population registry of Leipzig, Germany. Cross-sectional, standardised phenotyping included the assessment of various metabolic risk markers and ocular imaging, such as SD-OCT-derived thicknesses of ten optical bands of the retina. Global and Early Treatment Diabetic Retinopathy Study (ETDRS) subfield-specific optical retinal layer thicknesses were investigated in 7384 healthy eyes of 7384 participants from the LIFE-Adult-Study stratified by normal glucose tolerance, prediabetes (impaired fasting glucose and/or impaired glucose tolerance and/or HbA1c 5.7-6.4% [39-47 mmol/mol]) and diabetes. The association of optical retinal band characteristics with different indices of glucose tolerance (e.g. fasting glucose, area under the glucose curve), insulin resistance (e.g. HOMA2-IR, triglyceride glucose index), or insulin sensitivity (e.g. estimated glucose disposal rate [eGDR], Stumvoll metabolic clearance rate) was determined using multivariable linear regression analyses for the individual markers adjusted for age, sex and refraction. Various sensitivity analyses were performed to validate the observed findings. RESULTS: In the study cohort, nine out of ten optical bands of the retina showed significant sex- and glucose tolerance-dependent differences in band thicknesses. Multivariable linear regression analyses revealed a significant, independent, and inverse association between markers of glucose intolerance and insulin resistance (e.g. HOMA2-IR) with the thickness of the optical bands representing the anatomical retinal outer nuclear layer (ONL, standardised ß=-0.096; p<0.001 for HOMA2-IR) and myoid zone (MZ; ß=-0.096; p<0.001 for HOMA2-IR) of the photoreceptors. Conversely, markers of insulin sensitivity (e.g. eGDR) positively and independently associated with ONL (ß=0.090; p<0.001 for eGDR) and MZ (ß=0.133; p<0.001 for eGDR) band thicknesses. These global associations were confirmed in ETDRS subfield-specific analyses. Sensitivity analyses further validated our findings when physical activity, neuroanatomical cell/tissue types and ETDRS subfield categories were investigated after stratifying the cohort by glucose homeostasis. CONCLUSIONS/INTERPRETATION: An impaired glucose homeostasis associates with a thinning of the optical bands of retinal ONL and photoreceptor MZ. Changes in ONL and MZ thicknesses might predict early metabolic retinal alterations in diabetes.

Evaluation of a portable retinal imaging device: towards a comparative quantitative analysis for morphological measurements of retinal blood vessels.

This study investigated the possibility of using low-cost, handheld, retinal imaging devices for the automatic extraction of quantifiable measures of retinal blood vessels. Initially, the available handheld devices were compared using a Zeiss model eye incorporating a USAF resolution test chart to assess their optical properties. The only suitable camera of the five evaluated was the Horus DEC 200. This device was then subjected to a detailed evaluation in which images in human eyes taken from the handheld camera were compared in a quantitative analysis with those of the same eye from a Canon CR-DGi retinal desktop camera. We found that the Horus DEC 200 exhibited shortcomings in capturing images of human eyes by comparison with the Canon. More images were rejected as being unevaluable or suffering failures in automatic segmentation than with the Canon, and even after exclusion of affected images, the Horus yielded lower measurements of vessel density than the Canon. A number of issues affecting handheld cameras in general and some features of the Horus in particular have been identified that might contribute to the observed differences in performance. Some potential mitigations are discussed which might yield improvements in performance, thus potentially facilitating use of handheld retinal imaging devices for quantitative retinal microvascular measurements.

Automatic optic disc detection in colour fundus images by means of multispectral analysis and information content.

The optic disc (OD) in retinal fundus images is widely used as a reference in computer-based systems for the measurement of the severity of retinal disease. A number of algorithms have been published in the past 5 years to locate and measure the OD in digital fundus images. Our proposed algorithm, automatically: (i) uses the three channels (RGB) of the digital colour image to locate the region of interest (ROI) where the OD lies, (ii) measures the Shannon information content per channel in the ROI, to decide which channel is most appropriate for searching for the OD centre using the circular Hough transform. A series of evaluations were undertaken to test our hypothesis that using the three channels gives a better performance than a single channel. Three different databases were used for evaluation purposes with a total of 2,371 colour images giving a misdetection error of 3% in the localisation of the centre of the OD. We find that the area determined by our algorithm which assumes that the OD is circular, is similar to that found by other algorithms that detected the shape of the OD. Five metrics were measured for comparison with other recent studies. Combining the two databases where expert delineation of the OD is available (1,240 images), the average results for our multispectral algorithm are: TPR = 0.879, FPR = 0.003, Accuracy = 0.994, Overlap = 80.6% and Dice index = 0.878.

Retinal vascular changes in preterm infants: heart and lung diseases and plus disease.

PURPOSE: To report the retinal vascular features of preterm infants with congenital heart disease (CHD), lung disease (pulmonary hypertension [PH] and bronchopulmonary dysplasia [BPD]), and ROP with plus disease to determine whether these disease entities are distinguishable on the basis of retinal vessel morphology. METHODS: The medical records of preterm infants with CHD, lung disease, and ROP with plus disease were reviewed retrospectively. Qualitative vascular findings were validated using computer-based software to analyze 25 representative images, each corresponding to one infant's eye. The images were organized into five groups, based on clinical information. Vessel diameter (d) and tortuosity index (TI) were measured. RESULTS: A total of 106 infants (mean gestational age, 30.5 ± 2.22 weeks) were initially included. Ophthalmologic evaluation of preterm infants with CHD and lung diseases showed vascular tortuosity without vasodilation at the posterior pole as well as in the periphery. Quantitative analysis showed that venular diameter was significantly increased in the plus disease group (P = 0.0022) compared to other groups. There was significantly less tortuosity in both arterioles and venules in BPD (P < 0.001, P = 0.0453) compared with plus group. CONCLUSIONS: The patterns of retinal vascular tortuosity observed in preterm infants may be unique to different systemic congestive conditions and could have therapeutic implications.

Optic disc and macula detection in fundus images by means of template matching.

Various methods for detecting optic disc and macula in fundus images have been developed. Our aim is to propose a fairly easy method for detecting both features jointly. This is achieved by first correcting in homogenous luminosity using a polynomial approximation of the background of the images. Secondly, the use of the cross-correlation in the frequency domain between the images and a steerable template which contains both structures. The 38 photographs used in this work belong to a local database of patients suffering diabetic retinopathy along its four severity stages. Our results showed 100% optic disc centers located within the OD area and 90% macula centers located within the MC area.

Three-dimensional reconstruction of blood vessels extracted from retinal fundus images.

We present a 3D reconstruction of retinal blood vessel trees using two views of fundus images. The problem is addressed by using well known computer vision techniques which consider: 1) The recovery of camera-eyeball model parameters by an auto-calibration method. The camera parameters are found via the solution of simplified Kruppa equations, based on correspondences found by a LMedS optimisation correlation between pairs of eight different views. 2) The extraction of blood vessels and skeletons from two fundus images. 3) The matching of corresponding points of the two skeleton trees. The trees are previously labelled during the analysis of 2D binary images. Finally, 4) the lineal triangulation of matched correspondence points and the surface modelling via generalised cylinders using diameter measurements extracted from the 2D binary images. The method is nearly automatic and it is tested with 2 sets of 10 fundus retinal images, each one taken from different subjects. Results of 3D vein and artery trees reconstructions are shown.

NeuronGrowth, a software for automatic quantification of neurite and filopodial dynamics from time-lapse sequences of digital images.

We developed NeuronGrowth, a software for the automatic quantification of extension and retraction of neurites and filopodia, from time-lapse sequences of two-dimensional digital micrographs. NeuronGrowth requires a semiautomatic characterization of individual neurites in a reference frame, which is then used for automatic tracking and measurement of every neurite over the whole image sequence. Modules for sequence alignment, background subtraction, flat field correction, light normalization, and cropping have been integrated to improve the quality of the analysis. Moreover, NeuronGrowth incorporates a deconvolution filter that corrects the shadow-cast effect of differential interference contrast (DIC) images. NeuronGrowth was tested by analyzing the formation of outgrowth patterns by individual leech neurons cultured under two different conditions. Phase contrast images were obtained from neurons plated on CNS homogenates and DIC images were obtained from similar neurons plated on ganglion capsules as substrates. Filopodia were measured from fluorescent growth-cones of chick dorsal root ganglion cells. Quantitative data of neurite extension and retraction obtained by three different users applying NeuronGrowth and two other manually operated software packages were similar. However, NeuronGrowth required less user participation and had a better time performance when compared with the other software packages. NeuronGrowth may be used in general to quantify the dynamics of tubular structures such as blood vessels. NeuronGrowth is a free plug-in for the free software ImageJ and can be downloaded along with a user manual, a troubleshooting section and other information required for its use from http://www.ifc.unam.mx or http://www.ifc.unam.mx/ffm/index.html.

Parallel multiscale feature extraction and region growing: application in retinal blood vessel detection.

This paper presents a parallel implementation based on insight segmentation and registration toolkit for a multiscale feature extraction and region growing algorithm, applied to retinal blood vessels segmentation. This implementation is capable of achieving an accuracy (Ac) comparable to its serial counterpart (about 92%), but 8 to 10 times faster. In this paper, the Ac of this parallel implementation is evaluated by comparison with expert manual segmentation (obtained from public databases). On the other hand, its performance is compared with previous published serial implementations. Both these characteristics make this parallel implementation feasible for the analysis of a larger amount of high-resolution retinal images, achieving a faster and high-quality segmentation of retinal blood vessels.

Image analysis for retinopathy of prematurity diagnosis.

PURPOSE: To review findings from the authors' published studies involving telemedicine and image analysis for retinopathy of prematurity (ROP) diagnosis. METHODS: Twenty-two ROP experts interpreted a set of 34 wide-angle retinal images for presence of plus disease. For each image, a reference standard diagnosis was defined from expert consensus. A computer-based system was used to measure individual and linear combinations of image parameters for arteries and veins: integrated curvature (IC), diameter, and tortuosity index (TI). Sensitivity, specificity, and receiver operating characteristic areas under the curve (AUC) for plus disease diagnosis were determined for each expert. Sensitivity and specificity curves were calculated for the computer-based system by varying the diagnostic cutoffs for arterial IC and venous diameter. Individual vessels from the original 34 images were identified with particular diagnostic cutoffs, and combined into composite wide-angle images using graphics editing software. RESULTS: For plus disease diagnosis, expert sensitivity ranged from 0.308-1.000, specificity from 0.571-1.000, and AUC from 0.784 to 1.000. Among computer system parameters, one linear combination had AUC 0.967, which was greater than that of 18 of 22 (81.8%) experts. Composite computer-generated images were produced using the arterial IC and venous diameter values associated with 75% under-diagnosis of plus disease (ie, 25% sensitivity cutoff), 50% under-diagnosis of plus disease (ie, 50% sensitivity cutoff), and 25% under-diagnosis of plus disease (ie, 75% sensitivity cutoff). CONCLUSIONS: Computer-based image analysis has the potential to diagnose severe ROP with comparable or better accuracy than experts, and could provide added value to telemedicine systems. Future quantitative definitions of plus disease might improve diagnostic objectivity.

Comparison of the retinal microvasculature in European and African-Caribbean people with diabetes.

Diabetes aggravates the impact of elevated BP (blood pressure) on the microcirculation, and people of African ancestry with diabetes are more susceptible to microvascular damage than Europeans. In the present study, we investigated possible differences in the retinal microcirculation in people of European and African-Caribbean ethnicity with diabetes that might account for this. A total of 51 subjects with Type 2 diabetes (age 40-65 years; 25 male; 29 African-Caribbean) were studied. Clinic and 24 h ambulatory BP, and fasting glucose, insulin and lipids were measured. Digital retinal images were analysed using custom-written semi-automatic software to determine: LDR (length/diameter ratio) and AVR (arteriolar/venular diameter ratio), branching angles, vessel tortuosity and NT (number of terminal vessel branches). Arterioles were narrower in European people with diabetes than in African-Caribbean people with diabetes [mean (S.D.) arteriolar diameter, 76 (7) compared with 82 (11) microm respectively (P=0.03); arteriolar LDR, 28.1 (8.5) compared with 23.7 (7.0) respectively (P=0.046); and AVR, 0.66 (0.21) compared with 0.90 (0.36) respectively (P=0.028)]. Ethnic differences in arteriolar LDR, arteriolar diameter and AVR were not explained by differences in BP, but were attenuated by adjustment for the duration of diabetes. There was no significant relationship between BP and arteriolar narrowing in the group as a whole, although the relationship between arteriolar LDR and systolic BP was stronger in Europeans than African-Caribbeans [beta=0.08 (0.07) compared with beta=0.03 (0.06); P=0.03]. In conclusion, in the presence of diabetes, a relationship between BP and retinal arteriolar diameter was not evident and implies impaired small artery remodelling in the presence of diabetes. African-Caribbean people with diabetes have wider retinal arterioles and this could contribute to enhanced microvascular damage in this ethnic group.

Effect of antihypertensive treatment on retinal microvascular changes in hypertension.

OBJECTIVE: Hypertension causes arteriolar narrowing and rarefaction in the retinal circulation, but the extent to which these changes are reversible by antihypertensive treatment is not well studied. We compared the effect of antihypertensive treatment with a calcium-channel-blocker-based regimen and an angiotensin-converting-enzyme-inhibitor-based regimen on the retinal microvasculature. METHODS: Twenty-five patients (17 men, age range 24-71 years) with untreated hypertension were randomized to treatment with an amlodipine-based (n = 12) or lisinopril-based (n = 13) regimen in a double-blind, prospective parallel limb trial for 52 weeks. Measurements of blood pressure and the retinal microvasculature were made at baseline and at the end of the study. RESULTS: Both the amlodipine-based and lisinopril-based treatments reduced blood pressure to similar extents. Blood pressure reduction was associated with a reduction in arteriolar narrowing, a widening of arteriolar branch angle and an increase in arteriolar density. There were no significant differences between the two treatment regimens. CONCLUSION: Antihypertensive treatment is associated with improvement in arteriolar narrowing and rarefaction. Improved microvascular structure may contribute to the beneficial effects of antihypertensive treatment in hypertension.

Plus disease in retinopathy of prematurity: development of composite images by quantification of expert opinion.

PURPOSE: To demonstrate a methodology for generating composite wide-angle images of plus disease in retinopathy of prematurity (ROP), using quantitative analysis of expert opinions. METHODS: Thirty-four wide-angle retinal images were independently interpreted by 22 ROP experts as "plus" or "not plus." All images were processed by the computer-based Retinal Image multiScale Analysis (RISA) system to calculate two parameters: arterial integrated curvature (AIC) and venous diameter (VD). Using a reference standard defined by expert consensus, sensitivity and specificity curves were calculated by varying the diagnostic cutoffs for AIC and VD. From these curves, individual vessels from multiple images were identified with particular diagnostic cutoffs, and were combined into composite wide-angle images using graphics-editing software. RESULTS: The values associated with 75% underdiagnosis of true plus disease (i.e., 25% sensitivity cutoff) were AIC 0.061 and VD 4.272, the values associated with 50% underdiagnosis of true plus disease (i.e., a 50% sensitivity cutoff) were AIC 0.049 and VD 4.088, and the values associated with 25% underdiagnosis of true plus disease (i.e., 75% sensitivity cutoff) were AIC 0.042 and VD 3.795. Composite wide-angle images were generated by identifying and combining individual vessels with these characteristics. CONCLUSIONS: Computer-based image analysis permits quantification of retinal vascular features, and a spectrum of abnormalities is seen in ROP. Selection of appropriate vessels from multiple images can produce composite plus disease images corresponding to expert opinions. This method may be useful for educational purposes, and for development of future disease definitions based on objective, quantitative principles.

A method for semiautomatic tracing and morphological measuring of neurite outgrowth from DIC sequences.

We present a method for semi-automatic tracing and measuring of neurite outgrowth from time-lapse sequences of digital Nomarski micrographs. The algorithm is based on neurite ridge extraction and characterization from a single frame, followed by an automatic neurite tracking and measurement along the image sequence. Our method was tested with two sequences one containing 29 and other with 77 frames taken at intervals of 2 min. Our method rendered comparable length measurements but better time performance than measurements made by use of certain public software.

Retinal degeneration in children: dark adapted visual threshold and arteriolar diameter.

To assess the condition of the retina in children with retinal degeneration due to Bardet-Biedl syndrome (BBS, n=41), Leber congenital amaurosis (LCA, n=31), or Usher syndrome (USH, n=13), the dark adapted visual threshold (DAT) and arteriolar diameters were measured. Compared to controls, the initial DATs of nearly all (83/85) were significantly elevated, and in 26/62 with serial DATs, significant progressive elevation occurred. Arteriolar diameters were significantly attenuated and narrowed with age in BBS and USH, but not LCA. Higher DATs were associated with narrower arterioles. Such non-invasive procedures can document the natural history of these retinal diseases and have the potential to assess response to future treatment.

Evaluation of a computer-based system for plus disease diagnosis in retinopathy of prematurity.

OBJECTIVE: To measure accuracy and reliability of the computer-based Retinal Image Multiscale Analysis (RISA) system compared with those of recognized retinopathy of prematurity (ROP) experts, for plus disease diagnosis. DESIGN: Evaluation of diagnostic test or technology. PARTICIPANTS: Eleven recognized ROP experts and the RISA image analysis system interpreted a set of 20 wide-angle retinal photographs for presence of plus disease. METHODS: All experts used a secure Web site to review independently 20 images for presence of plus disease. Images were also analyzed by measuring individual computer-based system parameters (integrated curvature [IC], diameter, and tortuosity index) for arterioles and venules and by computing linear combinations and logical combinations of those parameters. Performance was compared with a reference standard, defined as the majority vote of experts. MAIN OUTCOME MEASURES: Diagnostic accuracy was measured by calculating sensitivity, specificity, and receiver operating characteristic area under the curve (AUC) for plus disease diagnosis by each expert, and by each computer-based system parameter, compared with the reference standard. Diagnostic agreement was measured by calculating the mean kappa value of each expert compared with all other experts and the mean kappa value of each computer-based system parameter compared with all experts. RESULTS: Among the 11 experts, sensitivity ranged from 0.167 to 1.000, specificity ranged from 0.714 to 1.000, AUC ranged from 0.798 to 1.000, and mean kappa compared with all other experts ranged from 0.288 to 0.689. Among individual computer system parameters, arteriolar IC had the highest diagnostic accuracy, with sensitivity of 1.000; specificity, 0.846; and AUC, 0.962. Arteriolar IC had the highest diagnostic agreement with experts, with a mean kappa value of 0.578. CONCLUSIONS: A computer-based image analysis system has the potential to perform comparably to recognized ROP experts for plus disease diagnosis.

Plus disease in retinopathy of prematurity: pilot study of computer-based and expert diagnosis.

PURPOSE: To measure accuracy of plus disease diagnosis by recognized experts in retinopathy of prematurity (ROP), and to conduct a pilot study examining performance of a computer-based image analysis system, Retinal Image multiScale Analysis (RISA). METHODS: Twenty-two ROP experts independently interpreted a set of 34 wide-angle retinal images for presence of plus disease. A reference standard diagnosis based on expert consensus was defined for each image. Images were analyzed by the computer-based system using individual and linear combinations of system parameters for arterioles and venules: integrated curvature (IC), diameter, and tortuosity index (TI). Sensitivity, specificity, and receiver operating characteristic areas under the curve (AUC) for plus disease diagnosis compared with the reference standard were determined for each expert, as well as for the computer-based system. RESULTS: Expert sensitivity ranged from 0.308 to 1.000, specificity ranged from 0.571 to 1.000, and AUC ranged from 0.784 to 1.000. Among individual computer system parameters, venular IC had highest AUC (0.853). Among all computer system parameters, the linear combination of arteriolar IC, arteriolar TI, venular IC, venular diameter, and venular TI had highest AUC (0.967), which was greater than that of 18 (81.8%) of 22 experts. CONCLUSIONS: Accuracy of ROP experts for plus disease diagnosis is imperfect. A computer-based image analysis system has potential to diagnose plus disease with high accuracy. Further research involving RISA system parameter cut-off values from this study are required to fully validate performance of this computer-based system compared with that of human experts.

Rod photoreceptor function predicts blood vessel abnormality in retinopathy of prematurity.

PURPOSE: To test the hypothesis that early rod dysfunction predicts the blood vessel abnormalities that are the clinical hallmark of retinopathy of prematurity (ROP). METHODS: Two rat models of ROP, induced by exposure to alternating 50%/10% oxygen (50/10 model) from postnatal day (P) 0 to P14, or exposure to 75% oxygen (75 model) from P7 to P14, and controls reared in room air were studied. In a longitudinal design, electroretinographic (ERG) records and digital fundus images were obtained at P20 +/- 1, P30 +/- 1, and P60 +/- 1. Rod sensitivity was derived from the ERG a-wave. Integrated curvature for the arterioles was calculated using Retinal Image multi-Scale Analysis (RISA) software. RESULTS: In both ROP models, rod sensitivity was low at P20. Sensitivity improved by P60 in the 50/10 model, but remained low in the 75 model. Integrated curvature was high at P20 in both ROP models, decreased nearly to normal by P30 in the 50/10 model, but remained high in the 75 model, even at P60. At P20, rod sensitivity correlated with integrated vessel curvature. Furthermore, low rod sensitivity at P20 predicted abnormal retinal vasculature--that is, high integrated curvature--at P30 and P60. In contrast, vessel curvature at P20 did not predict sensitivity at P30 or P60. CONCLUSIONS: The rods may instigate the vascular abnormalities that are the clinical hallmark of ROP.

Segmentation of blood vessels from red-free and fluorescein retinal images.

The morphology of the retinal blood vessels can be an important indicator for diseases like diabetes, hypertension and retinopathy of prematurity (ROP). Thus, the measurement of changes in morphology of arterioles and venules can be of diagnostic value. Here we present a method to automatically segment retinal blood vessels based upon multiscale feature extraction. This method overcomes the problem of variations in contrast inherent in these images by using the first and second spatial derivatives of the intensity image that gives information about vessel topology. This approach also enables the detection of blood vessels of different widths, lengths and orientations. The local maxima over scales of the magnitude of the gradient and the maximum principal curvature of the Hessian tensor are used in a multiple pass region growing procedure. The growth progressively segments the blood vessels using feature information together with spatial information. The algorithm is tested on red-free and fluorescein retinal images, taken from two local and two public databases. Comparison with first public database yields values of 75.05% true positive rate (TPR) and 4.38% false positive rate (FPR). Second database values are of 72.46% TPR and 3.45% FPR. Our results on both public databases were comparable in performance with other authors. However, we conclude that these values are not sensitive enough so as to evaluate the performance of vessel geometry detection. Therefore we propose a new approach that uses measurements of vessel diameters and branching angles as a validation criterion to compare our segmented images with those hand segmented from public databases. Comparisons made between both hand segmented images from public databases showed a large inter-subject variability on geometric values. A last evaluation was made comparing vessel geometric values obtained from our segmented images between red-free and fluorescein paired images with the latter as the "ground truth". Our results demonstrated that borders found by our method are less biased and follow more consistently the border of the vessel and therefore they yield more confident geometric values.

Plus disease in retinopathy of prematurity: an analysis of diagnostic performance.

PURPOSE: To measure agreement and accuracy of plus disease diagnosis among retinopathy of prematurity (ROP) experts; and to compare expert performance to that of a computer-based analysis system, Retinal Image multiScale Analysis. METHODS: Twenty-two recognized ROP experts independently interpreted a set of 34 wide-angle retinal photographs for presence of plus disease. Diagnostic agreement was analyzed. A reference standard was defined based on majority vote of experts. Images were analyzed using individual and linear combinations of computer-based system parameters for arterioles and venules: integrated curvature (IC), diameter, and tortuosity index (TI). Sensitivity, specificity, and receiver operating characteristic areas under the curve (AUC) for plus disease diagnosis were determined for each expert and for the computer-based system. RESULTS: Mean kappa statistic for each expert compared to all others was between 0 and 0.20 (slight agreement) in 1 expert (4.5%), 0.21 and 0.40 (fair agreement) in 3 experts (13.6%), 0.41 and 0.60 (moderate agreement) in 12 experts (54.5%), and 0.61 and 0.80 (substantial agreement) in 6 experts (27.3%). For the 22 experts, sensitivity compared to the reference standard ranged from 0.308 to 1.000, specificity from 0.571 to 1.000, and AUC from 0.784 to 1.000. Among individual computer system parameters compared to the reference standard, venular IC had highest AUC (0.853). Among linear combinations of parameters, the combination of arteriolar IC, arteriolar TI, venular IC, venular diameter, and venular TI had highest AUC (0.967). CONCLUSION: Agreement and accuracy of plus disease diagnosis among ROP experts are imperfect. A computer-based system has potential to perform with comparable or better accuracy than human experts, but further validation is required.