Purchase this article with an account.
S. Echegaray, G. Zamora, W. Luo, R. Kardon, J. Morales, P. Soliz; Automated Classification of Papilledema Using Frisen Grading and OCT Measurements. Invest. Ophthalmol. Vis. Sci. 2010;51(13):1775.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop and test a computer-based algorithm to assess digital images of the optic disc for signs of Papilledema. Based on reported observations by neuro-ophthalmologists, one of the earliest signs of disc edema is the change in sharpness or contrast especially in the superior and inferior regions of the disc margins. Our hypothesis is that region-based analysis of disc images allows classifying different stages of the disease.
A retrospective set of 28 retinal images, 1024x1024 pixels, 75% with papilledema, was used to train and test this algorithm. The ground truth consisted of the grading by three ophthalmologists using the Frisen scale for each image. The Average Total Retinal Thickness and the Average Retinal Nerve Fiber Layer Thickness measured by OCT were also provided. The algorithm extracts features from a number of regions of interest (ROIs) in each image. These ROIs were defined as annuli of pixels centered on the optic disc and four quadrants corresponding to superior, inferior, nasal, and temporal regions with respect to the optic disc. This allows the algorithm to detect quadrant-based intensity changes at different distances from the disc. The images were analyzed in three colorspaces, RGB, Lab, and YCBCr. Features include mean, variance, median, maximum change of intensity, and mean rate of intensity change (slope). A classification model was built by combining the features that best predicted OCT measurements individually. Leave-one-out cross correlation was used to test the performance of the model.
The best model for classifying the images into two classes, "normal-early" and "advanced", in terms of sensitivity (0.92), specificity (0.93), and area under the ROC curve (0.989), combined 28 features from the YCbCr color space that measure the speed of intensity changes away from the disc margins. The top 2 features were from the superior region, followed by two nasal. The top temporal region feature was ranked 27 in its contribution to the model.
These results show that correct classification of papilledema stage is possible using region-based features from optic disc images. This algorithm could be used for rapid screening of papilledema in clinical, intensive care, and emergency response settings.
This PDF is available to Subscribers Only