Abstract
Purpose::
Retinal blood vessels (RBV) and the fovea are useful features in scanning laser tomography (SLT) images. Currently the Heidelberg Retinal Tomographer (HRT) requires manual input to identify these features. The purpose of this study was to develop automated algorithms for BV and fovea detection for use with HRT SLT images.
Methods::
RBV Detection A ridge-based method was used, where ridges were defined as pixels where the image has an extremum in the direction of the largest surface curvature. Image ridges were extracted and, for each ridge pixel, a profile was obtained from adjacent pixels by sampling in the direction of largest surface curvature. Properties of the obtained profiles were used to define vectors that indicate RBV course. Vectors defining the properties of vessel segments (adjacent, similar ridge profiles) were then characterized to remove extraneous ridges. Post-processing of the ridge profile map extended and connected ridge branches, and removed isolated small branches. Performance was evaluated using the STARE database (IEEE MI, 2000, 19:203) of fundus photographs and compared to published vessel detection algorthms. Validation was achieved using HRT images of both normal and diabetic patients. Fovea Detection The minimum distance of each pixel to all neighboring ridges was computed using the post-processed RBV image. The center 128 x 128 pixels of the whole image were defined as the foveal candidate area. The pixel with the largest minimum distance value within this area was the estimate of the foveal location (fovea by distance transform, FDT). The foveal pit was searched for using a modified,morphological Top-Hat operator in the region adjacent to the FDT, giving a second estimate of the foveal location (fovea by Top-Hat, FTH). A series of decision rules selected either FDT or FTH as the best estimate. For 58 subjects the computed location of the fovea was compared with that identified manually.
Results::
When compared to published algorithms the RBV detection algorithm better identified RBV structure, particularly for abnormal data and low spatial resolution SLT images. The foveal detection algorithm was within 20 pixels of that identified manually for 44 of the 58 images analysed. 34 of these utilized the FTH estimate. 5 images were outside of 60 pixels (all FDT), 4 of which were eyes with retinopathy.
Conclusions::
Algorithms were developed that were capable of automated detection of the RBV architecture and fovea location for SLT images.
Keywords: image processing • retina • imaging/image analysis: clinical