Publication: Hybrid learning of vessel segmentation in retinal images
Issued Date
2021-01-01
Resource Type
ISSN
22869131
Other identifier(s)
2-s2.0-85096968030
Rights
Mahidol University
Rights Holder(s)
SCOPUS
Bibliographic Citation
ECTI Transactions on Computer and Information Technology. Vol.15, No.1 (2021), 1-11
Suggested Citation
Worapan Kusakunniran, Peeraphat Charoenpanich, Perapat Samunyanoraset, Sarocha Suksai, Sarattha Kanchanapreechakorn, Qiang Wu, Jian Zhang Hybrid learning of vessel segmentation in retinal images. ECTI Transactions on Computer and Information Technology. Vol.15, No.1 (2021), 1-11. doi:10.37936/ecti-cit.2021151.240050 Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/76757
Research Projects
Organizational Units
Authors
Journal Issue
Thesis
Title
Hybrid learning of vessel segmentation in retinal images
Other Contributor(s)
Abstract
In this paper, a novel technique of vessel segmentation in retinal images using a hybrid learning based approach is proposed. Unlike most other existing methods, a double-layer segmentation technique combining supervised and instance learning steps is introduced to enhance a sensitivity score of segmenting retinal blood vessels. The supervised learning based approach alone may not cope with unseen patterns caused by intrinsic variations in shapes, sizes, and color intensities of blood vessels across different retinal images. Thus, in the proposed hybrid learning solution, the supervised learning part is adopted to compute initial seeds of segmented vessels. They are then fed into the instance learning part as an initial foreground to further learn specific characteristics of vessels in each individual image. In the supervised learning step, the support vector machine (SVM) is applied on three types of features including green intensity, line operators, and Gabor filters. An iterative graph cut is adopted in the instance learning step, together with the pre-processing of morphological operations and the watershed algorithm. The proposed method is evaluated using two well-known datasets, DRIVE and STARE. It shows promising sensitivity scores of 82.6% and 82.0% on the DRIVE and STARE datasets respectively, and outperforms other existing methods in the literature.