Investigations on texture based feature extraction techniques for diabetic retinopathy detection

Abstract

In recent years, medical image processing plays a vital role in global health care for the benefit of the society. Eyes are the most vivacious organs which seeks immediate medical attention in case of any impairment. The prolonged diabetic patient may develop an eye abnormality, Diabetic Retinopathy (DR), causing damage to the retinal blood vessels. DR is initially asymptomatic but by the time an ignorant patient is diagnosed with DR, the severity gets worsen leading to visual impairment and blindness. Regular eye screening may prevent DR to attain proliferative stage, but it will waste expert s time in examining all the patients to identify the affected case 1 out of 100. Hence, there is a need to develop an automated system to detect DR from the fundus image at an early stage to save expert s time without manually segmenting the lesions which is really very cumbersome. DR detection incorporates two primary tasks namely feature extraction and classification. Since classification relies completely on feature extraction, this work investigates various feature extraction techniques. Compared with shape, color, intensity and morphological features, the texture features utterly describe the background of image providing discriminating capabilities for biomedical applications and so both the spatial and frequency domain method of texture based feature extraction techniques are explored. First, Multi-scale Amplitude Modulation-Frequency Modulation (AM-FM) technique, a frequency domain approach is employed to decompose the retinal image into instantaneous amplitude and frequency components by making use of directional Shearlet filterbank. newline