Abstract
Fractal dimension (FD) calculated from retinal vasculature is a biomarker which may aid in the diagnosis of the ocular and the systemic diseases at an early stage. Both multifractal and box-counting methods for calculating the FD are highly dependent on the segmentation accuracy of the images. In this paper, we propose the Fourier fractal dimension (FFD) method that does not have any such limitation as it eliminates the need for prior vascular segmentation. However, the robustness of FFD in differentiating the subgroups of retinal images must be investigated. In this paper two datasets: STARE and MESSIDOR are used in finding a significant difference between the subgroups of healthy and diseased retinal images. Furthermore, the effect of image pre-processing using the multi-scale self-quotient filtering (MSQ) and contrast-enhanced grey scale (CHGS) method on the FFD is also investigated. For the STARE dataset, the analysis is performed between normal and pathological as well as the healthy and background diabetic retinopathy images. For the MESSIDOR dataset, the FFD is calculated for normal, retinopathy grade 1 and grade 2 retinal images. The results show that the FFD can differentiate between two subgroups of retinal images. Moreover, for STARE dataset a significant difference between the healthy and background diabetic retinopathy images is found when the MSQ filtered images are used before calculating the FFD. This observation points to the fact that proper image pre-processing before the calculation of FFD can play a significant role in improving the robustness of the Fourier fractal analysis (FFA).
Subject Classification: