Exudates, retinal and statistical features detection from diabetic retinopathy and normal fundus images: An automated comparative approach

In this paper, an automated method is designed to extract important features from both normal and diabetic retinopathy (DR) retinal fundus images. The developed method comprises of four basic modules. The pre-processing module performs colour space conversion, Brightness Preserving Dynamic Fuzzy Histogram Equalization (BPDFHE) and De-correlation Stretching to enhance the contrast of the image based on pixel intensities. In retinal feature extraction module, Optic Disk is segmented by employing logical AND operation of significant bit planes. Retinal vessels are extracted using matched filter response, local entropy thresholding and length filtering techniques. For DR images, exudates are separately extracted using connected component analysis and optic disk elimination. Statistical features such as exudates area, kurtosis, entropy and Universal Image Quality Index (UIQI) are calculated and tabulated to show distinguishable differences of the features of normal images from that of the DR images. Finally a performance evaluation is carried out in terms of accuracy based on a comparison of exudates area detected by the ground truths and those detected by this automated method.