This invention relates to an automated, that is to say, a computer based system which is integrated with the current methods of image acquisition systems like fundus camera for early detection of diabetic retinopathy.
Diabetic Retinopathy (DR) Is a common cause of visual impairment among diabetic patients. Automated detection of exudates and microaneurysms In diabetic retinopathy can assist in early detection of disease and screening of large population. Early detection of disease is particularly important since in majority of the cases, the disease goes unnoticed until the later stages when it leads to partial/complete vision loss.
The automated DR detection system can assist in a big way in screening procedures since a large population has to be screened and that too repeatedly (about once a year). Such a system can significantly reduce:-
1) The need for specialist manpower.
2) The cost of screening
3) The screening time
Moreover, human evaluation is prone to errors. Digital color fundus images are used by opthalmologists to detect and track various diseases of the eye like diabetic retinopathy. The normal features of fundus Images are optic disk, blood vessels and fovea. The main features of diabetic retinopathy include exudates, microaneurysms and haemorrhages: Microaneurysms and exudates are the earliest signs of DR.
Microaneurysms are small saccular pouches caused by local distension of capillary walls. They appear as small red dots in the image, temporal to the ma'cula. The bigger blood clots are called haemorrhages and they appear in the later stages of the disease. Exudates are the yellow lipids in fundus Images with high intensity and contrast with background. Our algorithm can detect exudates under different illumination and color conditions. The bright circular region from where the blood vessels emanate is called the optic disk. The brighness, color and contrast of the optic disk are very similar to that of the exudates, Therefore the detection of the optic disk and its boundary, robustly and accurately is very important to avoid Its erroneous classification as an exudate. The fovea region defines the center of the retina and the area around fovea in a diameter of 5 mm (3.5 times optic disk diameter) is called macula. This region is responsible for detailed central vision such as reading.
This invention encompasses a robust and computationally efficient approach for the localization and segmentation of blood vessels, optic disk, fovea, exudates, microaneurysms and haemorrhages in retinal images. Blood vessels are segmented using morphological operations (closing) which exploit their contrast property with the brighter background. Optic disk is detected using the generic features of blood vessel intersection and high intensity properties. Exudates and microaneurysms are also extracted using morphological operations and then detected accurately using color models and generic lesion properties. Optic disk localization is a pre-requisite for avoiding false detection of exudates as both of them have similar brightness, color and contrast. The size, count and distribution of various lesions are used to indicate the severity of the disease.
The invention has the following advantages over the prior works in the following, among other salient aspects:
1. Fast method of optic disk detection using intersection of only the thicker blood vessels emanating from the optic disk. This is computationally efficient as the search space is reduced to just finding the intersection of thicker blood vessels rather than previous works which were focused on using the entire network.
2. Morphological operations are used together with color models and generic lesion properties to detect the microaneurysms, haemorrhages and exudates accurately. The use of these features makes our system robust against illumination changes and color variations.
3. A complete system for detecting various fundus image features very accurately and also to detect abnormalities and predict Diabetic Retinopathy.

The blood vessels are extracted using morphological closing operations. The green channel image is closed at 2 different sizes of a disk stnjcturing element. The result with the smaller structuring element is subtracted from the result with the larger structuring element, and thresholded to give the blood vessels. The larger structuring element size (radius) is chosen large enough (obtained by training) to close all vessels. The smaller structuring element size is varied to obtain vessel segmentation at different scales. In particular, the size of the smaller structuring element is chosen very close to the larger size to obtain just the thicker vessels in the fundus image. The size of the smaller structuring element is chosen very small compared to the larger structuring element size to obtain the entire vessel network in the fundus image.
Exudates appear as bright lesions in retinopathic images and have sharp edges and high contrast with the background. Most of the standard edge detectors add a lot of noise and miss out key edges when used for extracting exudate edges and hence are rot suitable for this application. We perform boundary detection for exudates using morphological operations. The green channel image is closed using structuring elements of 2 different large sizes and the results subtracted to extract the boundaries of the exudates candidate regions. The resulting image is thresholded and short breaks in contours are connected using smoothing splines. A morphological filling operation is performed to obtain the candidate regions (patches). However, the candidate regions may contain artifacts. Therefore, a linear classifrer is built which uses the brightness and edge properties of exudates. Exudates are bright yellow or white in color and have high intensity in the green channel. We localize the exudate patches more accurately by taking all the candidate regions whose mean intensities in the green channel are greater than a fraction (obtained by training the system) of the maximum intensity in the channel. For classifying the patches based on their edge strength, the gradient magnitude image of the green channel is chosen. This gradient magnitude image is thresholded (the absolute threshold obtained by training) and the number of white pixels in the thresholded image for each exudate patch is counted. Patches which do not have sufficient gradient count are discarded. Patches that satisfy both the brightness criterion and gradient count are retained and classified as exudates.
Optic disk invariably appears in exudates detection due to similarity in features. In order to remove the optic disk from the classifrcation result, we detect the optic disk using convergence of blood vessels and intensity property, The exudates result image is subtracted from the thicker blood vessels image and the result is passed through a morphological thinning operation. The vessel segments are then modeled as straight lines using Hough transform and only the lines with high slopes are retained as vessels arising from the optic disk have high slopes in the vicinity of the disk. The retained lines are intersected pairwlse to generate an intersection map (of intersection points). The intersection map and the high intensity feature of the optic disk are combined in a cost and the point which maximizes the cost in the fundus image is taken as the optic disk location, A disk mask is obtained by applying circular hough transform on the thresholded gradient image "(of green channel) in the vicinity of the disk location. Exudate patches appearing within the mask are removed. The thickest blood vessels arising from the optic disk and going above and below are referred to as the main blood vessels.
Fovea is detected by modeling the main blood vessels as a parabola and fovea is localized as the darkest pixel between 2 to 3 optical disk diameters along the main axis of the parabola. This Is because fovea is the point of low intensity in that region. A disk of 1 optic disk diameter around the fovea location is taken as the foveal region.
Microaneurysms and haemorrhages appear as dark patches on a brighter background. For the microaneurysms and haemorrhages (MAHMs) detection, a morphological filling operation is perfomned on the green channel and the unfilled image is subtracted from the filled image to obtain the candidate MAHMs patches. There are a lot of false patches and blood vessel segments in the MAHMs segmentation result. The blood vessels result having the entire vessel network is dilated and subtracted from the MAHMs result. We then use color model (obtained using blood vessels network) and remove all those patches which do not satisfy the color constraint. MAHMs need to have similar color (i.e. intensities in red, green channels) to nearby blood vessels. Fovea has similar properties to MAHMs and can appear

in the result. Hence any large dark patch appearing in the foveal region is discarded. The MAHMs patches are classified as either microaneurvsms or haemorrhaqes based on a size constraint.
The count and distribution of exudates and MAHMs is used to predict the severity of the disease. The medical standards set by the international council of ophthalmology (ICO) are used to characterize the severity of the disease. Accordingly, Diabetic Macular Edema (presence of exudates) is mild if there's retinal thickeninq or hard exudates in the posterior pole but distant from the macula. It Is Moderate if there are hard exudates approaching the center of the macula but not involving the center Finally it is severe if there are hard exudates involving the center of the macula. A person is free from Diabetic Macular edema if there are no hard exudates in the posterior pole. If there are no abnormalities, then the person Is free from diabetic retinopathy. If there are microaneurysms only, then it indicates the presence of mild non¬proliferative DR. If there are both microaneurysms and some small haemorrhages, then its moderate non¬proliferative DR. The DR is severe non-proliferative, proliferative if the MAHMs are larger In number and size as detailed in the ICO standards. The system for automated detection of the disease in the end displays the presence/absence of exudates and MAHMs and the level of severity of the disease.

We Claim:
1. Automated system for the early detection of diabetic retinopathy
comprising the steps of extraction of exudates which appear as bright patches with sharp edges in retinal images by morphological operations and further classification based on gradient strength and intensity properties; localizing the optic disk by combining the convergence of thicker blood arising from it along with high disk intensity property in a cost function along with blood vessel intersection, the blood vessel trunks themselves being obtained using morphological filters, the optic disk being masked out from the exudate detection to prevent false disease detections.; initially segmenting microaneurysms and hemorrhages (IVlAHMs) by morphological filters to exploit their local 'dark patch' property; developing a color model using blood vessels to further classify the segmented lesions and eliminate false ones; detecting and masking out the Fovea; determining the size, count and distribution of exudates and MAHMs, all the foregoing steps being carried out by ' computer screening and evaluation of the images of the retina.
2. Automated system for the early detection of diabetic retinopathy
substantially as herein described.