Abstract
Purpose: :
Our purpose is to develop a fast algorithm which segments the vessel tree on retina fundus images. Using the segmented images quantitative relevant measurements of the vessel tree such as arterial-venous ratio or vessel tortuosity can be efficiently gained to assist ophthalmologists.
Methods: :
The algorithm uses multiscale hierarchy to speed up the vessel segmentation of colored retina fundus images. The green channel of an image is rescaled to lower resolutions with a factor of 0.5 to compose a Laplacian resolution hierarchy. The full resolution image is used to extract the thin vessels and fine details, while the lower resolutions allow capturing thicker vessels. The vessel extraction on each resolution examines the eigenvalues of the Hessian matrix of the pixel’s neighborhood. The ratio of the largest and the smallest eigenvalue characterizes the vesselness of the examined pixels. The gained result images composed of these ratios are resized back to the original image size. The images are binarized using hysteresis threshold. The thresholded full resolution images are fused by union to obtain a binary vessel segmentation.
Results: :
The algorithm was tested on the STARE (20 test images) and on the DRIVE (20 test images) public eye fundus image databases which provide reliable hand segmentations as ground truth. On the STARE database an accuracy of 93.8% was gained with a specificity of 97.5% and a sensitivity of 65.1% to detect vessels. The average calculation time for 700x605 pixel images is 16 seconds. On the DRIVE database an accuracy of 94.9% was gained with a specificity of 96.8% and a sensitivity of 75.9% to detect vessels. The average calculation time for 565x584 pixel images is 11 seconds.
Conclusions: :
This method generates a competitive vessel segmentation. The proposed vessel detector applying different scales provides both effective results and low computation time. It can be used in the future to calculate vessel attributes such arterial-venous ratio for ophthalmologists.
Keywords: image processing • blood supply • imaging/image analysis: non-clinical