A Novel Approach for Detection of Optic Disc and Lesion Location for Screening Diabetic Retinopathy

Diabetic Retinopathy is not typically perceivable in diabetic patients at the initial stage. Their first signs, like micro-aneurysms, often go unnoticed in preliminary testing by specialists. Additionally, its presence is difficult to detect as there are other pathologies that may also lead to induce similar signs and symptoms. Until the detection of the presence of exudates, a specialist cannot simply deduce the presence of diabetic retinopathy. This paper presents a method to assist in the identification and differentiation of exudates on colour retinal images based on a variety of k-nearest neighbour filters. The proposed method proved to be a rational approach to detect bright lesions with sufficient certainty, yielding a possible injury with a specificity of 99%.

Diabetic Retinopathy Screening using Machine Learning for Hierarchical Classification

Diabetic Retinopathy is a consequence of prolonged unaddressed diabetes. It is currently diagnosed by the subjective clinical examination and manual grading of the fundus images by the ophthalmologists. This disease is progressive in nature. Hence early detection and treatment go a long way in helping the patients fight the dire consequences of the disease. Given the fact that number of ophthalmologists is very less as compared to the patients, a cost-effective, computer assisted, automated retina analysis system is highly desirable for the rural health care. This paper proposes an automatic Diabetic Retinopathy detection system based on the presence of bright lesions on the retina which is one of the symptoms of Diabetic Retinopathy. Initially the optic disc is removed from the fundus image as its brightness is similar to that of the bright lesions. Exudates are extracted and its various features are obtained. Later, feature based hierarchical classification is performed for detection of different stages of the disease. This method is based on the same logical steps as followed by the ophthalmologists and hence assures more accurate classification results. Two methodologies, Random Forest algorithm and Artificial Neural Network are explored and accuracy, sensitivity and specificity are evaluated at each stage of classification. The former outperformed the latter. The accuracy obtained using Random Forest are 100%, 85.71% and 87.5% and Artificial Neural network are 100%,78.5% and 66.67% for Stage 1, Stage 2 and Stage 3 respectively.

Segmentation and classification of bright lesions to diagnose diabetic retinopathy in retinal images

In view of predicting bright lesions such as hard exudates, cotton wool spots, and drusen in retinal images, three different segmentation techniques have been proposed and their effectiveness is compared with existing segmentation techniques. The benchmark images with annotations present in the structured analysis of the retina (STARE) database is considered for testing the proposed techniques. The proposed segmentation techniques such as region growing (RG), region growing with background correction (RGWBC), and adaptive region growing with background correction (ARGWBC) have been used, and the effectiveness of the algorithms is compared with existing fuzzy-based techniques. Images of eight categories of various annotations and 10 images in each category have been used to test the consistency of the proposed algorithms. Among the proposed techniques, ARGWBC has been identified to be the best method for segmenting the bright lesions based on its sensitivity, specificity, and accuracy. Fifteen different features are extracted from retinal images for the purpose of identification and classification of bright lesions. Feedforward backpropagation neural network (FFBPNN) and pattern recognition neural network (PRNN) are used for the classification of normal/abnormal images. Probabilistic neural network (PNN), radial basis exact fit (RBE), radial basis fewer neurons (RB), and FFBPNN are used for further bright lesion classification and achieve 100% accuracy.