Purchase this article with an account.
C. Spence, N. Gagvani, H. Li; Automated Detection of Neovascularization in Diabetic Retinopathy . Invest. Ophthalmol. Vis. Sci. 2004;45(13):2984.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose: Neovascularization is a pathognomatic sign of proliferative diabetic retinopathy. Urgent panretinal laser treatment is carried out when lesions are identified. However, small and early neovascularization is frequently under diagnosed, leading to laser treatment delay and decreased vision from developing sequential complications such as vitreous hemorrhage and/or traction retinal detachment. Effective computer aids could improve sensitivity and consistency of neovascularization detection during office visits or telemedicine consultations. More reliable detection would decrease the possibility of patients missing timely and effective laser treatment. Unlike microaneurysms, neovascularization shape and size varies, presenting extra challenges and requirements for automated detection systems. We applied a hierarchical pyramid/neural network (HPNN) to the complex issue of lesion detection. Methods: Digital 24–bit, 2,392 x 2,048 pixel fundus images with neovascularization (confirmed through fundus fluorescein angiograms) were selected to train the neural network system. Sub–images (256 x 256 pixel) of positive examples from 23 fundus images of 13 patients were extracted. Training was base on a "leave–one–out" procedure that trained an HPNN on all examples except those from one fundus (the "left–out") image. The HPNN was then tested on sub–images from the left–out fundus image. This was repeated for each fundus image, training and testing a separate HPNN for each. Digital images’ green color channel was chosen for classification due to its higher contrast than red or blue channels. Performance was measured from the combined test results. Results: The area under the receiver–operating–characteristic curve was 0.84 for distinguishing positive from negative subimages, and 0.79 for distinguishing fundus images with neovascularization from those without neovascularization. Conclusion: Trained HPNNs correctly classified fundus images without neovascularization. The application of computer–based techniques to automatically learn and detect complex object characteristics for use in ophthalmology is promising. Larger data sets and more thorough training (choice of regularization of constants, search of numbers of hidden units, etc.) should be explored.
This PDF is available to Subscribers Only