Purchase this article with an account.
Carla Agurto Rios, Honggang Yu, Victor Murray, Sheila C. Nemeth, Marios Pattichis, E Simon Barriga, Peter Soliz; Detection Of Hard Exudates In The Macula Using A Generalized Optimization Scheme. Invest. Ophthalmol. Vis. Sci. 2012;53(14):4100. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To present a system based on a generalized optimization scheme of image decompositions and classification to detect the presence of hard exudates in the macula of fundus images as means for automatic screening for potential clinically significant macular edema. This system can be extended to detect other bright and red lesions.
N=153 macula-centered digital fundus photographs were collected retrospectively from the University of Texas Health Science Center in San Antonio. The presence of lesions such as exudates, drusen, and microaneurysms were marked by a certified ophthalmic medical technologist. N=35 images were graded as normal, N=79 images presented hard exudates in the macula, and N=39 images presented only other lesions such as microaneurysms, hemorrhages, and drusen. The normal cases also contained images with retinal sheen and foveal reflex. We applied a 3-step algorithm for exudate detection on these images. First, possible candidate lesions are extracted from the amplitude-modulation frequency-modulation (AM-FM) decomposition of the images in different frequencies. Based on the output of these decompositions, binary maps that contained the candidates were selected through a process of optimization, where the optimal set of parameters was found using 53 images in the training stage. Second, features such as color, texture, and shape are obtained from each of the candidates extracted in the previous step. The objective of this step is to generate feature vectors that characterize the lesions and help us differentiate exudates from the rest of structures (e.g. foveal reflex, drusen,etc). Last, a Partial Least Squares classifier is trained with 53 images in order to classify the candidates on the testing set (100 images). After the candidates are classified, a threshold is chosen and the image classification rate is obtained.
The system achieved an AUC of 0.94 with best sensitivity/specificity of 92%/83% for candidate detection. By using the threshold set up on the previous step, the detection of maculas with exudates in a testing set of 100 images (51 with exudates, 49 without exudates) was 100% sensitive, with specificity of 58%.
A computer-aided detection algorithm based on generalized optimization scheme of image decompositions is presented. Given the process of optimization and the flexibility of the implementation this methodology could be applied to the detection of different type of lesions. In addition, the system does not require image enhancement since it is optimized for each image. The system achieves 100% sensitivity in detecting maculas with hard exudates. In future implementations, post-processing is going to be added in order to increase the specificity of the system.
This PDF is available to Subscribers Only