First Author: A.Ahmed UK
Co Author(s): R. Staff P. Gibbs S. Philip 0 0 0 0 0 0 0 0 0
Back to previous
347 million people worldwide had diabetes in 2012, the number continues to increase. Diabetic Retinopathy is the most common vascular complication and without early detection can lead to loss of vision. The development of retinal features such as blot haemorrhages and exudates around the macula are among signs clinically used to initiate further investigations; termed referable maculopathy. The automation of retinal and macular disease detection is becoming more important with the increase in diabetes, the importance of early detection in the management of the disease and the value of treatment response monitoring in both a clinical and drug development context.
Data was acquired from a national diabetic retinopathy screening service and analysed at the Aberdeen Biomedical Imaging Centre, University of Aberdeen. In this pilot study we used an established technique from other areas of medical imaging which has successfully been used in breast MRI to predict chemotherapy response prediction .
Two clinical patient groups were identified from the screening database; images of patients with known exudates (20) around the macula and no exudates around macula (10). These groups were defined using expert image evaluation and histology reports for each patient which we consider to be the gold standard. Here we chose to apply a commonly used technique; the spatial grey-level dependence matrix method proposed by Haralick. It is a statistical technique with the ability to study the 2nd order statistics of pixels at different spacings and angles. We used this texture analysis technique in an attempt to detect hamemorrhages by analysing an area within 1 disc diameter of the macula in retinal images of screening data. The software was written and applied to MR images previously  by author (A Ahmed) and was modified for use with retinal images. Using the matrix based programming language MatLab the software calculates 16 texture parameters within 1 disc diameter using a regions of interest around the macula. Here we aim to classify patients into their clinical groups defined above.
Using the spatial grey-level dependence matrix and 14 of Haralick’s texture equations with 2 additional features the software generates 16 texture parameters for each patient. Using a receiver operator characteristic (ROC) approach we analysed the performance of each parameter. That is we varied the threshold of significance and calculated the sensitivity and specificity at each threshold for the presences of exudates around the macular. Those that produced results significantly greater than chance are shown in Figure 1. The best performing parameter being cluster prominence with a sensitivity and specificity around 80%. Figure 1: shows Area under curve (ROC) values for texture parameters f4 (Variance), f7 (sum variance), f8 (sum entropy) and f16 (cluster prominence) as 0.735, 0.730. 0.740 and 0.765 respectively
Here we successfully used Texture Analysis to distinguish area around macula within 1 disc diameter between maculopathy and non-maculopathy categorised images. The performance of this approach was reasonable although inferior to some published reports, Flemings 2008 , Lee 2005  and Niemeijer 2005  who report sensitivity and specificity between 80-95% which have been optimised to particular samples. We are encouraged by these pilot results and aim to exploit the multiple parameters in order to improve our performance and examine the potential of the approach to detect imaging features such as microaneurysms and blot haemorrahages. We further aim to investigate the use of texture analysis as a tool for predicting disease associated with maculopathy and retinopathy.