skip to Main Content

Real-Time Optical Coherence Tomography Segmentation?

The recent publication, “Real-Time Automatic Segmentation of Optical Coherence Tomography Volume Data of the Macular Region” by Tian et al. [1] raises the obvious question of what actually constitutes “real-time” processing?

More broadly speaking, the paper reports on some segmentation results of OCT volumes acquired at the macula.  The method used is based on graph traversal techniques that, when applied to image segmentation, allow the developer to very easily define boundaries from one side of an image to another in terms of their lowest cost traversal.  For this traversal path to make sense, the input image data is first transformed into cost images that, ideally, have lowest pixel values at the boundaries of interest.  The transformations can be simple edge filtering, or more custom methods based on particular expected intensities or topographic constraints.  The resulting segmentation is not optimal in a global sense, and inherently 2d, but carries a low computational cost (O(n)).

Based on the efficiency of this technique, Tian et al. report the processing time of an average macula volume – 51 slices of 644-by-496 images – to be 26 seconds on a normal PC, a “2-8-fold increase in speed compared to other, similar reference algorithms used in the comparisons”.  Interestingly, in comparison to the method of Chiu et al [2], which uses the same graph-traversal approach, they achieved an impressive ten-fold speed increase based mainly on sound optimizations.  As an additional point of reference, on a similarly sized image volume, our Orion segmentation takes just 4 seconds.  As a further point of reference, the new multi-layer segmentation code for Heidelberg Engineering’s Spectralis instrument runs at around 2 seconds per B-scan; this means, it would take more than 100 seconds to process the aforementioned volume.

The accuracy reported on normal eyes with respect to manually delineated boundaries was very good, having an average unsigned error of just a pixel.  Consequently, the software, OCTRIMA 3D (OCT Retinal IMage Analysis 3D), as developed by Professor DeBuc’s group at the Bascom Palmer Eye Institute at the University of Miami, is stated as being “a leap forward towards achieving reliable, real-time analysis of 3D OCT retinal data.”

The work should be readily applauded: the method is strong and pragmatic, has a decent mechanism for extendibility to 3d, and their comparison to other algorithms thorough, allowing the reader to well understand how the reported technology truly fares.  That said, the title is misleading, which brings us back to the initial question of what constitutes real-time.

Real-time processing usually means the ability to do something at least 30 times a second; i.e., the execution time should be around 33ms.  This would require that the Bascom Palmer algorithm needs to run ~780 times faster to operate in real-time.  Admittedly, the implementation is in Matlab, so could be improved just via a porting to optimized C++.  Matlab, however, does allow for compiled, optimized code and also allows for code execution to be run on GPUs, so it is not clear what level of optimization had been achieved.

Fortunately, Voxeleron offer exactly such technologies today.  Namely, available for immediate licensing, we offer frame-rate 2D OCT segmentation of the human retina.  Although this is not approved for clinical use, development is based on our internal IEC 62034 compliancy, it could be used as an adjunct tool in surgical devices or also for automatic patient positioning during image acquisition.  Our 3D OCT segmentation is not quite at frame-rate – we’re working on that – but it is still the fastest, validated algorithm available [3].

[1] Tian J, Varga B, Somfai GM, Lee WH, Smiddy WE, Cabrera DeBuc D.  Real-Time Automatic Segmentation of Optical Coherence Tomography Volume Data of the Macular Region.  PLoS One. 2015 Aug 10;10(8).

[2] Chiu SJ, Li XT, Nicholas P, Toth CA, Izatt JA, Farsiu S.   Automatic segmentation of seven retinal layers in SDOCT images congruent with expert manual segmentation.  Optics express. 2010;18(18):19413–19428.

[3] Oakley JD, Gabilondo I, Songster C, Russakoff DB, Green A, Villoslada P.  Assessing Manual versus Automated Segmentation of the Macula using Optical Coherence Tomography.  IOVS, ARVO Meeting Abstracts, 2014.

Back To Top