A new supervised retinal vessel segmentation method based on robust hybrid features. (September 2016)

• Record Type: Journal Article
• Authors: Aslani, Shahab; Sarnel, Haldun
• Is Part Of: Biomedical signal processing and control. Volume 30(2016)
• Page Start: 1
• Page End: 12
• Publication Date: 2016-09
• Subjects: Retinal blood vessels -- Segmentation -- Hybrid feature vector -- Random forest; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2016.05.006
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Robust features of each pixel gathered to construct a hybrid feature vector. B-COSFIRE filter response was utilized as a feature in the supervised method. The result of segmentation in STARE database is superior to all state-of-art methods. The result of cross training is superior to all state-of-art methods. The result of segmentation in pathology images is superior to all state-of-art methods. Abstract: In this paper, we propose a new supervised retinal blood vessel segmentation method that combines a set of very robust features from different algorithms into a hybrid feature vector for pixel characterization. This 17-D feature vector consists of 13 Gabor filter responses computed at different configurations, contrast enhanced intensity, morphological top-hat transformed intensity, vesselness measure, and B-COSFIRE filter response. A random forest classifier, known for its speed, simplicity, and information fusion capability, is trained with the hybrid feature vector. The chosen combination of the different types of individually strong features results in increased local information with better discrimination for vessel and non-vessel pixels in both healthy and pathological retinal images. The proposed method is evaluated in detail on two publicly available databases DRIVE and STARE. Average classification accuracies of 0.9513 and 0.9605 on the DRIVE and STARE datasets, respectively, are achieved. When the majority of the common performance metrics are considered, our method is superior to the state-of-the-art methods. Performance results show that our method also outperforms the state-of-the-art methods in both cross training and pathological cases.