Purpose
Abnormalities in retinal vasculature provide useful information about clinical and sub-clinical cerebrovascular, cardiovascular, and metabolic health of the patient. This is especially true for diabetic retinopathy (DR), a common microvascular complication of diabetes in which damaged retinal blood vessels become leaky or occluded, ultimately leading to visual loss. Despite its promise, current clinical practice does not include retinal vessel analysis for DR primarily because the analysis involves highly laborious task of extracting vessels. We present a segmentation method that can automatically mark retinal vasculature and thus aid in automated DR screening.
Methods
Our vasculature segmentation method utilizes novel and customized image analysis techniques including image normalization, multi-scale putative vessel detection, and multi-orientation morphological filterbank analysis for vessel estimation.<br /> We evaluate our segmentation method on the publicly available DRIVE and STARE retinal image datasets with vasculature marked in great detail. We compute the accuracy and the area under the receiver operating characteristic (AUROC) at pixel-level for our segmentation method using all the images in each dataset.
Results
Vessel segmentation maps generated by our vasculature segmentation method for three retinal images are shown in Figure 1. Our method is robust to the varying pixel intensity ranges and qualities of the retinal images. Our method achieves accuracy of 95.3% and AUROC of 0.932 on the DRIVE dataset and accuracy of 95.6% and AUROC of 0.914 on the STARE dataset. These results are equivalent to the performance of a human grader.
Conclusions
We present a new approach for retinal vasculature segmentation that has the potential for real-world use with its high accuracy, invariance to imaging conditions, and good generalizability.