Enhancing retinal blood vessel segmentation in medical images using combined segmentation modes extracted by DBSCAN and morphological reconstruction. (August 2021)

• Record Type: Journal Article
• Authors: Mardani, Kamran; Maghooli, Keivan
• Is Part Of: Biomedical signal processing and control. Volume 69(2021)
• Publication Date: 2021-08
• Subjects: DBSCAN -- Morphological reconstruction -- Pattern recognition -- Retinal blood vessel segmentation -- Noise cancellation; 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.2021.102837
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: DBSCAN has been used to segment the vessels in medical images. Several thin and thick vessels are found and a larger pattern of the vessel tree is obtained. This method can be used for complex Tubular-shape images and low-quality old images. Abstract: In this study, a new algorithm is proposed for retinal blood vessel segmentation in medical images using the density-based spatial clustering of applications with noise (DBSCAN) and the morphological reconstruction (MR). The images are thus divided into smaller equal rectangular ones, called cells through the R variable. Upon finding the best DBSCAN clustering parameters, neighborhood radius ( ɛ ), and number of minimum points ( Z ), the pattern of the retinal blood vessels appears in the cells. Noise removal operations are also performed via median, logical AND operator ( && ), and MR filters, respectively. The MR is accordingly used in a loop with an incremental K counter, as a square structuring element. Finally, the given vessels are segmented by setting four parameters of R, ɛ, Z, and K . After combining different Z and K modes, the resulting images of the thin veins are joined to the thick ones, and most of the gaps in the vessel tree are corrected. Several veins, not visible to the naked eye in the original images, are also detected. The proposed method is implemented for three datasets DRIVE, STARE and the VAMPIRE. Performance measures including sensitivity ( Se ) and area under the curve (AUC) correspondingly show a significant increase in these datasets. Accuracy ( Acc ) and specificity ( Sp ) values are further comparable to the results by others, and a slight decrease in some cases is attributed to the new vessels found in addition to the ground truth (GT) images.