Abstract
Purpose::
Fundus photography is a non-invasive technique for in vivo ophthalmoscopic inspection of retinal disorders. Quantitative information about the vascular network can facilitate clinical diagnosis of retinal diseases. We propose an algorithm for segmentation of the vascular network by using multi-scale analysis in selected wavelet channel frames.
Methods::
We use the STAR database for testing our segmentation algorithm consisting of twenty datasets. The images are captured by a TopCon TRV-50 fundus camera with 35 degree field of view. An over-complete wavelet frame expansion is performed. We perform selective channel rejection in the decomposition tree followed by wavelet shrinkage and enhancement operators to separate retinal objects from background. The obtained vessel likelihood map is the basis for a Bayesian classifier into vessel and non-vessel employing shape, topology, and intensity cues.
Results::
The level set segmentations were compared to the expert grading on a pixel-by-pixel basis. The mean sensitivity and specificity were 0.88 ± 0.09 and 0.98 ± 0.02.
Conclusions::
Multi-scale wavelet frame is an ideal domain for retinal vessel segmentation and analysis. Future work is devoted to increase segmentation performance of small scale vasculature.
Keywords: imaging/image analysis: clinical