Purchase this article with an account.
Pascal A. Dufour, Ute E. Wolf-Schnurrbusch, Lala Ceklic, Sebastian Wolf, Jens H. Kowal; A Quick Pathology Hinting System for Optical Coherence Tomography Data. Invest. Ophthalmol. Vis. Sci. 2012;53(14):3113.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
With the fast advancements in Optical Coherence Tomography (OCT), the amount of image data that has to be inspected by the clinician is ever increasing. New tools are needed to allow a quick assessment of a patients slice stack and find possibly pathological areas without the need to screen through every individual B-scan.We built a hinting system that is able to automatically analyze OCT volume scans and find abnormally thick or thin cell layers to focus the attention of the ophthalmologist to potentially pathological areas.
The underlying algorithm of the hinting system is an automatic segmentation of six cell layer boundaries, coupled with an active shape model.We built an active shape model from 60 segmentations from healthy eyes. This model is able to deform to any healthy shape of the retina and it can easily be deformed to fit a segmentation of a new dataset. However, because we did not use pathological datasets for the training, the model is not able to represent any form of morphological abnormality.Given any patients OCT slice stack, the active shape model is deformed to best fit its segmentation. The fit will be good for healthy regions but bad in the areas of morphological abnormalities. The distance of the active shape model to the segmentation serves as a measure of confidence of how healthy the retina at any position is.The ability of the active shape model to represent the whole range of healthy shapes was evaluated with a leave-one-out test.Performance analysis was carried out using 15 OCT volumes of patients with drusen to evaluate time and memory requirements and measure the accuracy of the hinting.
The leave-one-out test of the active shape model resulted in a mean unsigned error of 5.7±5.4 μm, demonstrating a good fit to healthy segmentations.The required execution time was less than 15 seconds and the required memory was below 1GB.Deviations of more than 15 μm from the normal cell layer thickness were detected with high confidence.
The proposed hinting system is both fast and has low hardware requirements, making it suitable for a clinical setup. It is a useful tool to cope with the ever-increasing amounts of data and is especially valuable for finding abnormalities at very early stages in disease progression.
This PDF is available to Subscribers Only