Purchase this article with an account.
Julie Cho, Amir H Kashani, Mark S Humayun; Spectral Classification of Retinal Features Using K-Means Clustering Algorithm . Invest. Ophthalmol. Vis. Sci. 2015;56(7 ):5261.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop effective methods of sampling and classifying retinal tissue spectra obtained with a hyperspectral computed tomographic imaging spectrometer.
High-resolution hyperspectral images (Figure 1a) from three species (rabbits, canine and minipig) were acquired using a hyperspectral computed tomographic imaging spectrometer (HCTIS) that has been previously described in detail (Kashani AH et al., 2011, 2012, 2013). Animals were sedated and imaging was performed under general anesthesia through pharmacologically dilated pupils using an HCTIS mounted on a Zeiss FF450 fundus camera. The HCTIS records up to 76 spectral bands from 450-700nm within a single snapshot obtained by a standard fundus photograph. For this study, multiple images of characteristic regions from each animal were obtained including arteries, veins, retina and optic disc. Spectra from multiple regions of interest within individual images, across images, and across species were manually inspected for similar spectral characteristics. Spectra were also quantitatively analyzed with k-means clustering algorithm and Voronoi diagram was generated (Figure 1d).
Spectral characteristics of retinal vessels and tissue are robust and largely correlated to vascularity (Figure 1b-d) and pigmentation in the animal models. Vascular structures were reliably identified across multiple species and required minimal dimensionality. Three-dimensional data were best clustered with differing k-values (k=3 dog, k=2 pig, k=1 rabbit). Features such as retinal pigmentation across species could be determined by characteristic heterogeneity of each tissue as determined by standard deviation of the mean (Figure 1e). Characterization of these structures required multi-dimensional models, but segmentation was possible with silhouette means above 0.6 in the case k=1 and above 0.8 in case k=2 and k=3 indicating compactness of the data (Figure 1e).
Characterization of retinal features in multiple spectral dimensions allows robust and reliable classification of retinal tissue. Here we sampled multiple images and found defining spectral characteristics of structural features within and across images as well as across species. These results suggest that clustering methods can be used to help classify tissue types in hyperspectral images.
This PDF is available to Subscribers Only