Purchase this article with an account.
Gwen Musial, Hope M Queener, Suman Adhikari, Hanieh Mirhajianmoghadam, Alexander W Schill, Nimesh Bhikhu Patel, Jason Porter; Automatic segmentation of retinal capillaries in adaptive optics perfusion images using a convolutional neural network. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1496. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Metrics of retinal capillary structure may be useful biomarkers for retinal and optic nerve head (ONH) diseases. Despite their high resolution, adaptive optics (AO) perfusion images can possess large variations in contrast, intensity, and background noise, thereby limiting the use of global or adaptive thresholding methods. We sought to develop an automated segmentation approach to quantify perfused capillaries in AO images.
Perfusion images were generated by calculating the standard deviation of each pixel’s intensity over time from stabilized split detector AO videos taken 2-6° from the ONH when focused at the nerve fiber layer. A convolutional neural network (CNN) was trained to classify 4 pixel classes: capillary, larger vessel, background, image border. Manually marked capillaries were < 20 pixels (~20 μm) in diameter. To train the CNN, 9,112 patches (128 x 128 pixels) were extracted from 150 Gaussian filtered (σ = 3) AO images (758 x 758 pixels) from 4 normal human and 6 normal non-human primate eyes. To test the CNN, 1,936 patches were extracted from 12 Gaussian filtered AO images from 3 eyes that were not part of the training set. The CNN’s probability map was converted to 4 class maps by Otsu’s method. A second rater marked all test images. To account for slight shifts between the same marked capillaries (Fig. 1b, green arrow), we developed a modified Dice coefficient to include segmentations that were separated by less than the average capillary radius (5 pixels) as true positives.
CNN capillary segmentation had an accuracy of 0.938. Mean traditional and modified Dice coefficients for the test set were 0.621 and 0.908 between rater 1 and CNN, and 0.613 and 0.910 between raters 1 and 2. Mean densities (%) across the test set were similar for rater 1, 2, and CNN (11.4 ± 2.3, 11.9 ± 2.6, 11.2 ± 3.4; P=.84).
Close agreement between the CNN and manual segmentations may enable the robust estimation of perfused capillary metrics. The algorithm separates capillaries from diffuse background noise, distinguishes capillaries that cross larger vessels (Fig. 1b, orange arrow), and may translate to other imaging modalities.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
Figure 1. Representative manual and automatic segmentations: (a) AO image of perfused vasculature. (b) Image from (a) showing manual (red) and CNN (blue) capillary segmentations, and regions of overlap (purple). Scale bar: 50μm.
This PDF is available to Subscribers Only