An Improved Retinal Vessel Segmentation Framework Using Frangi Filter Coupled With the Probabilistic Patch Based Denoiser

Vessel segmentation has come a long way in terms of matching the experts at detection accuracy, yet there is potential for further improvement. In this regard, the accurate detection of vessels is generally more challenging due to the high variations in vessel contrast, width, and the observed noise level. Most vessel segmentation strategies utilize contrast enhancement as a preprocessing step, which has an inherent tendency to aggravate the noise and therefore, impede accurate vessel detection. To alleviate this problem, we propose to use the state-of-the-art Probabilistic Patch-Based (PPB) denoiser within the framework of an unsupervised retinal vessel segmentation strategy based on the Frangi filter. The PPB denoiser helps preserve vascular structure while effectively dealing with the amplified noise. Also, the modified Frangi filter is evaluated separately for tiny and large vessels, followed by individual segmentation and linear recombination of the binarized outputs. This way, the performance of the modified Frangi filter is significantly enhanced. The performance evaluation of the proposed method is evaluated on two recognized open-access datasets, viz: DRIVE and STARE. The proposed strategy yields competitive results for both preprocessing modalities, i.e., Contrast Limited Adaptive Histogram Equalization (CLAHE) and Generalized Linear Model (GLM). The performance observed for CLAHE over DRIVE and STARE datasets is (<inline-formula> <tex-math notation="LaTeX">$Sn = 0.8027$ </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">$Acc = 0.9561$ </tex-math></inline-formula>) and (<inline-formula> <tex-math notation="LaTeX">$Sn = 0.798$ </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">$Acc = 0.9561$ </tex-math></inline-formula>), respectively. For GLM, it is observed to be (<inline-formula> <tex-math notation="LaTeX">$Sn = 0.7907$ </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">$Acc = 0.9603$ </tex-math></inline-formula>) and (<inline-formula> <tex-math notation="LaTeX">$Sn = 0.7860$ </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">$Acc = 0.9583$ </tex-math></inline-formula>) over DRIVE and STARE datasets, respectively. Furthermore, based on the conducted comparative study, it is established that the proposed method outperforms various notable vessel segmentation methods available in the existing literature.