A Novel Blood Vessel Extraction Using Multiscale Matched Filters with Local Features and Adaptive Thresholding

Retinal blood vessel structure is an important feature for computer-aided diagnosis and treatment of diseases including diabetic retinopathy, hypertension, glaucoma, obesity, arteriosclerosis and retinal artery occlusion, and an accurate extraction is required to improve the accuracy of the diagnostic task. This paper proposes a new algorithm for blood vessel segmentation and extraction in retinal images. A multiscale matched filter combined with local features is developed to effectively extract blood vessels from retinal images. Local features are extracted from a circular and adaptive window around a candidate blood vessel pixel. Experimental evaluation using publicly available DRIVE and STARE databases shows accurate extraction of vessel networks as demonstrated by improved false alarm rates and segmentation accuracy when compared against existing works. The mean true positive rate (TPR) values obtained are (0.7661%) and (0.6312 %) for STARE and DRIVE datasets respectively, while the mean false positive rate (FPR) values achieved are (0.0311 %) for STARE and (0.0183 %) for DRIVE. Moreover, our proposed method gave high accuracy values when compared to similar work on same datasets, 93.53% and 94.73% for STARE and DRIVE datasets respectively. While in the cases of methods achieving higher accuracy value than ours, we either have a higher TPR or a lower FPR. These promising results can be enhanced in the future by deploying some other features and/or experimenting different thresholding techniques. In addition to the detection of blood vessels from retinal images, there is ongoing work to develop a quantitative method based on the shape and regularity of the blood vessels detected in order to detect possible signs / symptoms of a disease.