Automated Identification of Diabetic Retinal Exudates using Support Vector Machines and Neural NetworksRichard Markham, Alireza Osareh, Majid Mirmehdi, Barry Thomas, Maria Macipe, Automated Identification of Diabetic Retinal Exudates using Support Vector Machines and Neural Networks. The Association for Research in Vision and Ophthalmology Conference. May 2003. No electronic version available.
Purpose: To develop a system for automated identification of exudates in digital retinal photographs of diabetic retinopathy.
Methods: We used 142 colour retinal images obtained from a non-mydriatic retinal camera, but with papillary dilation, using a 45 degree field of view. Image resolution was 760 x 570 at 24 bit RGB. Two pre-processing steps, normalisation of colour images to a pre-selected standard, and local contrast enhancement distributing the value of the pixels around the local mean to facilitate later segmentation, were performed.
Candidate exudate regions were identified by carrying out a two stage colour segmentation algorithm based on a Gaussian-smoothed histogram analysis and Fuzzy C-Means clustering. Fuzzy approaches provide a mechanism to represent and manipulate uncertainty and ambiguity in the identification of fundus lesions, allowing pixels to belong to multiple classes with varying degree of membership.
In order to distinguish exudates from other lesions of similar appearance it is necessary to select a set of significant distinguishing features. Colour spaces which separate luminance from chrominance are most suitable. We used mean Luv and standard deviation of the Luv values inside a candidate region, mean Luv and standard deviation of the Luv values surrounding the region, Luv values of the region centroid, region size, region compactness and region edge strength.
Classification of the identified regions into exudates and non-exudates was performed using either Support Vector Machines or Neural Networks. A Neural Network based approach using a Back Propagation learning method (NN-BP) performed marginally better than the a support vector machine method, but the latter was more flexible in controlling the trade off between sensitivity and specificity rates.
Results: The NN-BP based system could achieve 95.5% sensitivity and 73.3% specificity for the identification of patients with exudate retinopathy by classifying whole images. Detection rats could be even higher for individual exudates within an image but at the cost of reduced specificity overall. We looked to see what types of mistakes (false positive or false negative) were made by these systems compared with the clinical judgement of an ophthalmologist. The commonest problem, reducing specificity, was that small light reflexes were often recorded as exudates.
Conclusions: Accurate automated identification of exudates is possible. Further development of this system should allow the production of a clinically useful system.