Feature Of The Week 11/8/12: University Of Bern Presents Methods for Multi-Surface Segmentation of OCT Images
Over the past several years and for the foreseeable future, development and implementation of algorithms for processing of OCT images has become a fertile area for researchers at universities and product development engineers at OCT system companies. Such algorithms can dramatically improve OCT images and extract new information, both of which can improve clinical decision making and add tremendous value and differentiation to a product line. Topical areas include removing motion artifacts from 3D data sets, fast processing and rendering of very large data sets, extracting functional information, registration of data sets from longitudinal studies, and more. Recently the University of Bern has published some interesting work on “Graph-Based Multi-Surface Segmentation of OCT Data Using Trained Hard and Soft Constraints” and provided their software for researchers. Below is a summary of their recent work.
Segmentation of Optical Coherence Tomography (OCT) datasets is an important prerequisite for the analysis of the retina. Thickness measurements of the retinal cell layers provide valuable information about the retinal structure. It provides quantitative measurements that can be used for diagnostic purposes, clinical studies and trials, and research into disease progression. As manual segmentation is extremely time-consuming, automatic methods are required.
We recently developed a graph-based method that is capable of segmenting OCT datasets of healthy eyes as well as drusen in datasets of patients with Age-Related Macular Degeneration (AMD). The advantage of graph-based segmentation methods is that the global optimal solution is found. This is coupled with the ability to segment multiple surfaces simultaneously.
The improvement over previous methods is the inclusion of additional soft constraints. On one hand, soft smoothness constraints increase the rigidity of a segmented surface, making it less susceptible to outliers such as vessel shadows and decrease the influence of noise. On the other hand, soft constraints were added in-between neighboring surfaces and act as forces that drive the solution towards a specified model. The parameters of these soft constraints are learned from training datasets and vary depending on the position in the image. In that way, true local information is incorporated into the segmentation algorithm.
Of interest to researchers is that we made our developed application publicly available. It currently supports the segmentation of OCT datasets from Heidelberg Spectralis machines and automatically segments six layers in healthy eyes and eyes of patients with AMD. Manual correction of the automatic segmentation or manual segmentation from scratch is also supported. Measurement capabilities include the standard ETDRS sectors, area and point measurements in B-scans, and average reflectivity measurements of cell layers.
For more information see Article. Courtesy Pascal A. Dufour from University of Bern.