Abstract
Purpose: :
To develop and validate an automated system to analyze digital fundus images for staging and monitoring swelling of the optic disc due to raised intracranial pressure.
Methods: :
A retrospective set of 294 longitudinal, digital fundus photographs of the optic disc from 39 subjects with papilledema was used. Ground truth was determined from independent analysis by three neuro-ophthalmologists using the modified Frisen Scale. A suite of image analysis algorithms was developed to quantify the features used by human expert analysis: 1) blurriness of the optic disc border was assessed by analyzing the location along radial lines from the center of the disc with the largest change in pixel intensity, 2) texture features of the RNFL were quantified within a peripapillary retinal annulus, 3) vessel obscuration was defined by a new metric, the vessel continuity index (VCI), that quantifies the continuity along each vessel’s radial length. We used a decision tree forest algorithm for classification and leave-one-out cross correlation to test the performance of our model.
Results: :
The algorithm showed substantial agreement (Κ = .71, p < 0.001) with ground truth when grading papilledema per patient, and substantial agreement (Κ = .61, p < 0.001) when evaluating per image. Each of the image features that were quantified changed with grade of papilledema.
Conclusions: :
These results show that it is feasible to automatically detect and stage papilledema. The algorithm provides objective, quantitative, accurate, and repeatable assessment of the stage of papilledema at levels of accuracy comparable to those of expertly trained neuro-ophthalmologists. This algorithm could be used for rapid analysis of digital images acquired in clinical, intensive care, and emergency response settings by non-ophthalmologists to diagnose papilledema and its severity.
Keywords: optic disc • optic nerve • image processing