Purchase this article with an account.
Pascal Dufour, Hannan Abdillahi, Lala Ceklic, Ute Wolf-Schnurrbusch, Sebastian Wolf, Jens Kowal; Automatic Segmentation of OCT Datasets Using Superpixels-Classification. Invest. Ophthalmol. Vis. Sci. 2013;54(15):5509.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Retinal cell layer segmentation of Optical Coherence Tomography (OCT) data is an important prerequisite for diagnostic purposes, clinical trials and research into pathologies of the eye. Most methods apply a layer-based segmentation scheme and try to detect cell layer boundaries. However, when the retinal cell layers are disrupted, these methods are incapable of providing a meaningful segmentation. To overcome that limitation, we propose a method that does not focus on finding the cell layer boundaries, but instead assigns each pixel in a B-scan to a retinal layer.
Using the Simple Linear Iterative Clustering (SLIC) algorithm, each B-scan is first divided into superpixels; small regions with homogeneous texture that can be assumed to lie on the same retinal cell layer. Various image filters are then applied in the proximity of each superpixel, providing a set of features descriptive of the neighborhood of that superpixel. Using these features, a trained Random Forest classifier computes the probability of each superpixel belonging to a specific retinal cell layer. Finally, a multi-label graph-cut algorithm aggregates the results from the classification. In this work, 1225 B-scans from 25 OCT volumes of healthy eyes were used to train the Random Forest. Each volume was first automatically segmented and then manually inspected and corrected. The following 7 regions in the volume were segmented: vitreous, NFL, combined GCL and IPL, combined INL and OPL, ONL, layers between IS/OS and BM, and choroid.
20 OCT volumes from 20 healthy subjects were used to evaluate the algorithm. From each volume, 5 B-scans were randomly chosen and manually segmented by two observers. The Dice coefficient was computed to test the accuracy of the automatic segmentation. A mean Dice coefficient of 0.945±0.0094 over all regions and B-scans was achieved, which is comparable to the inter-observer variability of a Dice coefficient of 0.947±0.0081.
The proposed method can distinguish the retinal layers in normal eyes with high accuracy, without assuming continuous cell layers. A next step will be the training of the classifier on OCT data with advanced-stage pathologies, where a layer-based segmentation cannot be applied. Initial tests on OCT data of patients with diagnosed Wet Age-Related Macular Degeneration show promising results.
This PDF is available to Subscribers Only