10. Pathologies Segmentation in Eye Fundus Images Based on Frequency Domain Filters

This chapter presents a novel method based on discrete frequency transforms to segment various pathologies in eye fundus color images such as exudates, blood vessels, and aneurysms. Non-uniform illuminated eye fundus images are corrected by applying a homomorphic high-pass frequency filter. Then, a super-Gaussian band-pass filter defined in the frequency transform domain is used to distinguish between background and foreground objects. The filtering step works with the green channel that usually contains the most relevant information to segment different pathologies. Specifically, exudates detection after transform inversion of the filtered image requires a gamma correction to enhance foreground objects. Otsu’s thresholding method is applied to the enhanced image and masked over the effective area to get the segmented exudates. For blood vessels and aneurysms, back in the spatial domain, the negative of the filtered image is required. Then a median filter is applied to reduce noise or artifacts followed by gamma contrast enhancement. Again, Otsu’s thresholding method is used for image binarization. Next a morphological closing operation is applied and masking the effective image area gives the segmented blood vessels or aneurysms. Illustrative examples using retinographies from a free public domain clinical database are included to demonstrate the capability of the frequency filter approach.