A practical approach to OCT based classification of Diabetic Macular Edema

This paper addresses the problem of automatic classification of OCT images for identification of patients with DME versus normal subjects. In this paper a relativity simple and practical approach is proposed to exploit the information in OCT images for a robust classification of Diabetic Macular Edema (DME) using coherent tensors. From the retinal OCT scan top and bottom layers are extracted to find thickness profile. Cyst spaces are also segmented out from the normal and DME images. The features extracted from thickness profile and cyst are tested on Duke Dataset having 55 diseased and 53 normal OCT scans. Results reveal that SVM with Leave-one-Out gives the maximum accuracy of 79.65% with 7.6 standard deviation. However, experiments reveal that for the identification of DME, nearly same accuracy of 78.7% can be achieved by using a simple threshold which can be calculated using thickness variation of OCT layers. Moreover a comparison of the proposed algorithm on a standard dataset with other recently published work shows that our method gives the best classification performance.