Dataset Information

Automatic Segmentation of Retinal Capillaries in Adaptive Optics Scanning Laser Ophthalmoscope Perfusion Images Using a Convolutional Neural Network.

ABSTRACT: Purpose:Adaptive optics scanning laser ophthalmoscope (AOSLO) capillary perfusion images can possess large variations in contrast, intensity, and background signal, thereby limiting the use of global or adaptive thresholding techniques for automatic segmentation. We sought to develop an automated approach to segment perfused capillaries in AOSLO images. Methods:12,979 image patches were extracted from manually segmented AOSLO montages from 14 eyes and used to train a convolutional neural network (CNN) that classified pixels as capillaries, large vessels, background, or image canvas. 1764 patches were extracted from AOSLO montages of four separate subjects, and were segmented manually by two raters (ground truth) and automatically by the CNN, an Otsu's approach, and a Frangi approach. A modified Dice coefficient was created to account for slight spatial differences between the same manually and CNN-segmented capillaries. Results:CNN capillary segmentation had an accuracy (0.94), a Dice coefficient (0.67), and a modified Dice coefficient (0.90) that were significantly higher than other automated approaches (P < 0.05). There were no significant differences in capillary density and mean segment length between manual ground-truth and CNN segmentations (P > 0.05). Conclusions:Close agreement between the CNN and manual segmentations enables robust and objective quantification of perfused capillary metrics. The developed CNN is time and computationally efficient, and distinguishes capillaries from areas containing diffuse background signal and larger underlying vessels. Translational Relevance:This automatic segmentation algorithm greatly increases the efficiency of quantifying AOSLO capillary perfusion images.

SUBMITTER: Musial G

PROVIDER: S-EPMC7424955 | biostudies-literature | 2020 Jul

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Automatic Segmentation of Retinal Capillaries in Adaptive Optics Scanning Laser Ophthalmoscope Perfusion Images Using a Convolutional Neural Network.

Musial Gwen G Queener Hope M HM Adhikari Suman S Mirhajianmoghadam Hanieh H Schill Alexander W AW Patel Nimesh B NB Porter Jason J

Translational vision science & technology 20200723 2

<h4>Purpose</h4>Adaptive optics scanning laser ophthalmoscope (AOSLO) capillary perfusion images can possess large variations in contrast, intensity, and background signal, thereby limiting the use of global or adaptive thresholding techniques for automatic segmentation. We sought to develop an automated approach to segment perfused capillaries in AOSLO images.<h4>Methods</h4>12,979 image patches were extracted from manually segmented AOSLO montages from 14 eyes and used to train a convolutional ...[more]

PMID: 32855847

Similar Datasets

Project description:To develop and test the application of an adaptive optics scanning laser ophthalmoscope (AOSLO) with eye tracking for high-resolution microperimetric testing.An AOSLO was used to conduct simultaneous high-resolution retinal imaging and visual function testing in six normal subjects. Visual sensitivity was measured at test locations between the fovea and 5.0° eccentricity via an increment threshold approach using a 40-trial, yes-no adaptive Bayesian staircase procedure (QUEST). A high-speed eye tracking algorithm enabled real-time video stabilization and the delivery of diffraction-limited Goldmann I-sized stimuli (diameter = 6.5 arc min = ?32 ?m; ? = 680 nm) to targeted retinal loci for 200 ms. Test locations were selected either manually by the examiner or automatically using Fourier-based image registration. Cone spacing was assessed at each test location and sensitivity was plotted against retinal eccentricity. Finally, a 4.2 arc min stimulus was used to probe the angioscotoma associated with a blood vessel located at 2.5° eccentricity.Visual sensitivity decreases with eccentricity at a rate of -1.32 dB/deg (R = 0.60). The vertical and horizontal errors of the targeted stimulus delivery algorithm averaged 0.81 and 0.89 arc min (?4 ?m), respectively. Based on a predetermined exclusion criterion, the stimulus was successfully delivered to its targeted location in 90.1% of all trials. Automated recovery of test locations afforded the repeat testing of the same set of cones over a period of 3 months. Thresholds measured over a parafoveal blood vessel were 1.96 times higher (p < 0.05; one-tailed t-test) than those measured in directly adjacent retina.AOSLO-based microperimetry has the potential to test visual sensitivity with fine retinotopic precision. Automated recovery of previously tested locations allows these measures to be tracked longitudinally. This approach can be implemented by researchers interested in establishing the functional correlates of photoreceptor mosaic structure in patients with retinal disease.

Project description:Fundus albipunctatus (FA) is a form of congenital stationary night blindness characterized by yellow-white spots, which were classically described as subretinal. Although night blindness and delayed dark adaptation are hallmarks of this condition, recent studies have described a macular phenotype, particularly among older patients. Using a fluorescence adaptive optics scanning laser ophthalmoscope (FAOSLO), this study provides in vivo morphologic data at the cellular level in FA.To study the cone photoreceptors and the albipunctate spots in FA at single-cell resolution.A woman in her 30s with FA underwent a complete ophthalmic examination, including conventional imaging tests, at the University of Rochester. A FAOSLO was used to obtain infrared reflectance images of the cone mosaic at the central fovea and along the superior and temporal meridians to 10° eccentricity. Cone density was measured at the foveal center, and cone spacing was calculated in sampling windows eccentrically. In the area of the albipunctate spots, autofluorescence FAOSLO images (excitation, 561 nm; emission, 624 ? 40 nm) were simultaneously obtained.Structural appearance of cones, cone density and spacing, and reflectance and autofluorescence of albipunctate spots.Cone density was reduced to 70% of the lower limit of the normal range at the foveal center (78.7 × 10(3) cones/mm(2); mean [SD] reference range, 199 [87] × 10(3) cones/mm(2)), and cone spacing was increased eccentrically to 10° (sign test, P =?.045). Individual cone central core reflectances appeared dim, suggesting loss of photoreceptor outer segments. The albipunctate spots were hypoautofluorescent. No photoreceptors or retinal pigment epithelium cells were identified at the locations of the albipunctate spots.Although the predominant clinical symptom of night blindness and the electroretinography results suggest a primary rod dysfunction, examination with a FAOSLO demonstrates that cone density is also reduced. This finding may represent an early sign of progression to macular phenotype in FA. The hypoautofluorescence suggests that the albipunctate spots do not represent lipofuscin.

Project description:PurposeDual wavelength retinal oximetry has been developed for adults, but is not available for infants. Retinal oximetry may provide insight into the pathophysiology of oxygen-mediated diseases like retinopathy of prematurity. More insight in the oxygen metabolism of the retina in infants may provide valuable clues for better understanding and subsequent prevention or treatment of the disease. The measurements of oxygen saturation are obtained with two fundus images simultaneously captured in two different wavelengths of light. The comparison in light absorption of oxygenated and deoxygenated hemoglobin can be used to estimate the oxygen saturation within the retinal vessels by means of a software algorithm. This study aims to make retinal oximetry available for neonates. The first step towards estimating retinal oxygen saturation is determining the optical density ratio. Therefore, the purpose of this study is to image healthy newborn infants with a scanning laser ophthalmoscope and determine the optical density ratio for retinal oximetry analysis.MethodsImages of the retina of full-term healthy infants were obtained with an SLO, Optomap 200Tx (Optos), with two laser wavelengths (532nm and 633nm). The infant lay face down on the lower arm of the parent, while the parent supported the chest and chin with one hand, and stabilized the back with the other hand. No mydriatics or eyelid specula were used during this study. The images were analyzed with modified Oxymap Analyzer software for calculation of the Optical Density Ratio (ODR) and vessel width. The ODR is inversely and approximately linearly related to the oxygen saturation. Measurements were included from the superotemporal vessel pair. A paired t-test was used for statistical analysis.ResultsFifty-nine infants, (58% female), were included with mean gestational age of 40 ± 1.3 weeks (mean ± SD) and mean post-natal age of 16 ± 4.8 days. A total of 28 images were selected for retinal oximetry analysis. The ODR was 0.256 ± 0.041 for the arterioles and 0.421 ± 0.089 for the venules (n = 28, p < 0.001). The measured vessel-width for the arterioles was 14.1 ± 2.7 pixels and for the venules 19.7 ± 3.7 pixels (n = 28, p < 0.001).ConclusionsRetinal oximetry can be performed in newborn infants by combining an SLO and a dual-wavelength algorithm software. Sensitivity of the approach is indicated by the fact that the ODR measurements are significantly different between the arterioles and the venules. However, more variability in ODR is seen with the SLO approach in babies than is seen with conventional oximetry in adults. This approach is completely non-invasive, non-contact and even avoids the use of mydriatics or eyelid specula.

Dataset Information

Automatic Segmentation of Retinal Capillaries in Adaptive Optics Scanning Laser Ophthalmoscope Perfusion Images Using a Convolutional Neural Network.

Publications

Automatic Segmentation of Retinal Capillaries in Adaptive Optics Scanning Laser Ophthalmoscope Perfusion Images Using a Convolutional Neural Network.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets