Purchase this article with an account.
Chaithanya Ramachandra, Sandeep Bhat, Muneeswar G Nittala, Srinivas R Sadda, Kaushal Solanki; EyeMark: Automated Image-based Biomarker Computation Tools for Monitoring Diabetic Retinopathy. Invest. Ophthalmol. Vis. Sci. 2014;55(13):4827.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Microaneurysm (MA) turnover rates (appearance and disappearance rate of MAs) are a good predictor of likelihood of progression to more severe retinopathy, establishing MA turnover as an excellent biomarker for diabetic retinopathy. Measuring this quantity from color fundus images involves two labor intensive steps: careful alignment of current and baseline images, and marking of individual MAs. This process is very time consuming and prone to error, if done entirely by human graders. We propose an automated MA turnover computation tool - EyeMark.
100 longitudinal retinal fundus color image sets with images in each set collected at least 6 months apart were obtained retrospectively from University of Southern California Keck School of Medicine. The MAs in these images were annotated by an expert grader. The longitudinal images were aligned using condition number theory based registration in the vesselness transform domain. Region of interest (ROI) was determined based on morphological filtering and Hessian matrix based computations to limit the pixels that are processed further. Using the ground truth annotation in a 50/50 train/test split, pixels were classified as MA/non-MA using novel multi-scale descriptors such as Gaussian derivatives and morphological filterbanks in a support vector machine framework. Classified pixels are grouped into blobs to determine the MAs in a given image. The MA locations in longitudinal images are matched to determine number of persistent MAs and to compute turnover rates.
All longitudinal image pairs of fair/adequate quality were successfully registered even with multiple lesion changes, and different intensities between the image pair (Figure 1). MA detection has high sensitivity with AUROC of 0.92 at lesion/blob level. The turnover range (Figure 2) computed across longitudinal images was in good agreement with the turnover computed using the ground truth annotations.
Automated MA turnover computation requires two key challenges to be resolved - robust longitudinal image registration and accurate MA detection. With EyeMark we have tackled these issues and achieved accurate MA turnover computation.
This PDF is available to Subscribers Only