Purchase this article with an account.
Michael Henry Goldbaum, Nima Hatami; Accurate Retinal Artery and Vein Classification using Local Binary Patterns. Invest. Ophthalmol. Vis. Sci. 2014;55(13):232.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
This paper proposes a new Local Binary Pattern (LBP)-based method to generate features for automatic artery and vein (AV) classification of retinal fundus images.
Ten retinal fundus images were selected from the STARE dataset for validating the proposed method. Each image was digitized to produce a 1400x1200 pixel image, 24-bits per pixel (standard RGB). All ten images contained abnormalities that obscured or confused the blood vessel appearance. Segmented vessels were divided into pieces. LBP windows were applied along each point on a vessel centerline in a gray scale image. Generated features were classified by Random Forest Trees (RFT).
Figure 1 shows the LBP operation feature generation on a fundus image. Figure 2 shows the step-by-step pre-processing (Matched Filter Response and morphological operation) for vessel center-line extraction and the final AV classification results obtained using the LBP features for RFT classifiers. The result of the automatic AV classifier matches well the hand-labeled image by two medical experts. Area under the ROC curve was 0.95 over all vessel pieces in all images.
LBP classified by RFT is a fast and accurate method for AV classification of color fundus images. The proposed method performs well in the image margins compared to existing methods and is also illumination- and scale-invariant. This simple method is robust, even when applied to low contrast and quality fundus images. LBP accuracy compares favorably to previous AV classification methods. Accurate AV classification enables future applications, such as AV diameter ratio (AVR) and artery and venous tortuosity measurements.
This PDF is available to Subscribers Only