Purchase this article with an account.
E. S. Barriga, S. R. Russell, M. S. Pattichis, V. M. Murray, S. Murillo, H. T. Davis, III, M. D. Abramoff, P. Soliz; Relationship Between Visual Features and Analytically Derived Features in Non-exudated AMD Phenotypes: Closing the Sematic Gap. Invest. Ophthalmol. Vis. Sci. 2009;50(13):3274.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To evaluate a new image processing technique for assigning phenotypes to digital images of the retina for patients with non-exudative (NE) age-related macular degeneration (AMD) and to determine the relationship of these computer-based features to visually described age-related macular degeneration (AMD) phenotypes.
The dataset consisted of 100 digital images from 100 sequential patients with NE AMD. Images were scanned from 35mm slides at 2400x2400 pixels and centered on the fovea. This data set was previously used to produce 12 phenotypes by the ophthalmologists, e.g. temporal drusen, small distributed drusen, geographic atrophy, etc. A process for decomposing an image into its frequency and amplitude components was applied and used to cluster the images into sets with similar mathematical characteristics. To test the validity of the computer-generated feature-based model of phenotypes, an experiment was conducted by having the ophthalmologists and 2 other trained analysts evaluate the visual similarity of images from the same cluster. Twenty experiments were conducted whereby a target image was selected randomly. The four closest images according to their analytically derived features were extracted, along with one outlier (random image outside the target’s cluster). The analysts were asked to select the outlier and grade the remaining four images as close or not close in similarity.
Nine feature scales representing different frequency and amplitude bands were selected. A total of 150 features representing different frequency scales and amplitudes were used to characterize the retinal images. The average correct selection of the outlier was 70% (p<.001). When asked to select the next least like the target image, the correct selection of the outlier was 85%.
These results suggest that certain features, such as those that represent the frequency and amplitude characteristics of an image, can be correlated to visual features as seen by the human observer. These results suggest that visually recognized fundus patterns (fundus phenotypes) of NE AMD can be detected and defined using computer vision-based methods. Amplitude-modulation and frequency-modulation techniques demonstrate a high accuracy in sorting fundus patterns into phenotypes based on the presentation of the disease.
This PDF is available to Subscribers Only