Skip to content

Voxeleron Announces Release of a New Standard in OCT Retinal Segmentation for Ophthalmic OCT Vendors

The Septem algorithm accurately segments seven retinal layers in seconds

Pleasanton, CAOctober 6, 2012 – Voxeleron LLC announces the immediate availability of vendor independent software for the automatic segmentation of seven retinal layers in optical coherence tomography (OCT) volumes.  The Septem algorithm is based on Voxeleron’s proprietary denoising methods, applicable to both high and low-resolution OCT scans.

Jonathan Oakley, Co-Founder and Principal Scientist, says that their methods facilitate faster and more robust analysis of existing volumetric scan patterns and pave the way for lower-cost devices to offer the same analyses as the high-end devices of today.

“Noise in OCT data is an inherent barrier to accurate and repeatable analyses.  We found that by being able to more accurately resolve the true signal in the data, we were able to segment more layers and do so with higher accuracy and even better repeatability.  It sounds obvious, but we only truly saw this by focusing our efforts on scans of lower signal quality or scans with disruptive pathologies.  The ability to segment seven retinal layers reliably, quickly, and over a wide range of signal conditions makes the algorithm truly unique.  We have seen significantly lower error rates in comparison to commercially available inner retina segmentation algorithms, where fewer layers are segmented.  To achieve this level of performance requires years of development time.  And we were only able to do so based on our extensive experience in the development and release of image processing algorithms for ophthalmic OCT.”

7 Layer Segmentation

An early collaboration with Johns Hopkins School of Medicine showed very promising results.

“The development of retinal segmentation software that works across many OCT platforms is a major advance for the field,” says Dr. Peter Calabresi, Professor of Neurology at Johns Hopkins.  “The quantitative measures of the inner and outer nuclear layers appear to have predictive value for estimating clinical disability in multiple sclerosis.”

Robert Chang, M.D., Assistant Professor of Ophthalmology at the Stanford School of Medicine, provided additional clinical input and scan data to support the development of the algorithm.

“Based on an early study of their results, I was encouraged to see the Voxeleron algorithm accurately segment layers on scans where a commercially available algorithm had failed,” says Dr. Chang. “With the added ability to segment more retinal layers, I am interested in looking at new parameters for disease. We looked at numerous eyes and saw consistent results, and we plan to publish our findings.”

This table is based on a manual review of the segmentation results of 192 eyes:

Algorithm No Failure Minor Failure Major Failure
Cirrus 6.0 154 12 26
Septem 175 13 4
  1. No failure: no error was seen in the segmentation result.
  2. Minor failure: errors were seen, but not within the measurement annulus (this annulus is used in the Cirrus 6.0 software and described in Mwanza et al.)
  3. Major failure: clear errors were seen that would affect the measurement taken within the annulus.

The data came from the Byers Eye Institute at Stanford School of Medicine

Comparison Of Thickness Maps
Example thickness maps from the Stanford review.

The two thickness maps shown above were taken from the Stanford review.  On the left is the thickness map produced by Cirrus 6.0 software (Carl Zeiss Meditec Inc., Dublin, CA).  On the right, the results from the Septem software.  The Zeiss algorithm has erroneously identified two very thin arcs in the inferior of the scan.  The annulus is also erroneously thick on the nasal side (this is the right eye), as the RNFL has been incorrectly segmented there.

There is a growing need for better accuracy, repeatability, and robustness to meet the ever-evolving demands of clinical applications.  Perhaps the best example is in the field of neuro-ophthalmology where quantitative measurements of the inner retinal layers are of tremendous interest.  A seminal study by Saidha et al. found a unique subset of multiple sclerosis (MS) patients with primary retinal pathology showing disproportionate thinning of the inner and outer nuclear layers, and more rapid disease progression, suggesting a direct pathologic process in retinal neurons of these patients that may be associated with cortical atrophy.

As endpoints for clinical trials, OCT and inner-retinal segmentation algorithms offer unique insight into a number of neurological pathologies.  In a different publication they concluded:  “As such, ganglion cell layer plus inner plexiform layer thickness as measured by retinal segmentation of optical coherence tomography scans may be an ideal marker for monitoring neurodegeneration within the eye and may provide a feasible, specific primary outcome measure for evaluating neuroprotective agents in clinical trials.”

More recently, the same Johns Hopkins-led collaboration was able to show that retinal measures do reflect global central nervous system pathology in MS, “with thicknesses of discrete retinal layers each appearing to be associated with distinct central nervous system processes.”

Such advancement of scientific understanding is based on quantitative segmentation tools being utilized  by expert clinicians.  The software used in these studies was unable, however, to delineate the interface between the inner nuclear and outer plexiform layers, and the outer plexiform and IS/OS junction.  Septem’s ability to provide such additional information allows, for example, the inner nuclear layer (INL) to be measured in isolation.  This has the potential to further the understanding, perhaps leading  to more specific trials and endpoints.  Just this week, Saida et al. concluded, “Increased INL thickness on OCT is associated with disease activity in MS. If this finding is confirmed, INL thickness could be a useful predictor of disease progression in patients with MS.”

Image Showing An Eye With A Staphyloma, Accurately Segmented By The Septem Algorithm.
Image showing an eye with a staphyloma, accurately segmented by the Septem algorithm.

In the ophthalmic market, as demand for lower-cost, smaller, and more mobile imaging devices increases, image quality will inevitably be compromised.  The Septem algorithm is ideally suited to maintain the accuracy of quantification methods in the face of lower quality images.  Conversely, as the data rates of high-end devices increases, fast and reliable algorithms become even more essential.  The clinical relevance of wider field of view scan patterns with OCT has already been reported — further justification for the collection of more data and an algorithm that performs well across the entire posterior pole.  Given stiff competition and rapid progress in OCT technology along these dimensions, the next generation of retinal segmentation algorithms will need to be fast, light, generally applicable, and easily adaptable to different images, layers, and scan patterns.  Taking just 3 seconds on an off-the-shelf PC to accurately segment seven layers in an OCT volume comprised of 67 million points, Voxeleron’s Septem algorithm is all of these things and moreInquiries regarding the licensing of this software can be directed to

Onh Example
An example of the Septem algorithm run on an Optic Nerve Head scan pattern. Further analysis of this algorithm is planned by the end of the year.

References Mentioned in this Post

Profile and predictors of normal ganglion cell-inner plexiform layer thickness measured with frequency-domain optical coherence tomography.  Mwanza JC, Durbin MK, Budenz DL, Girkin CA, Leung CK, Liebmann JM, Peace JH, Werner JS, Wollstein G; Cirrus OCT Normative Database Study Group.   Invest Ophthalmol Vis Sci. 2011 Oct 4;52(11):7872-9. Print 2011 Oct.

Primary retinal pathology in multiple sclerosis as detected by optical coherence tomography.  Saidha S, Syc SB, Ibrahim MA, Eckstein C, Warner CV, Farrell SK, Oakley JD, Durbin MK, Meyer SA, Balcer LJ, Frohman EM, Rosenzweig JM, Newsome SD, Ratchford JN, Nguyen QD, Calabresi PA.  Brain. 2011 Feb;134(Pt 2):518-33. Epub 2011 Jan 20.

Relationships Between Retinal Axonal and Neuronal Measures and Global Central Nervous System Pathology in Multiple Sclerosis.  Saidha S, Sotirchos ES, Oh J, Syc SB, Seigo MA, Shiee N, Eckstein C, Durbin MK, Oakley JD, Meyer SA, Frohman TC, Newsome S, Ratchford JN, Balcer LJ, Pham DL, Crainiceanu CM, Frohman EM, Reich DS, Calabresi PA.  Arch Neurol. 2012 Oct 1:1-10. doi: 10.1001/archneurol.2013.573.

Optical coherence tomography segmentation reveals ganglion cell layer pathology after optic neuritis.  Syc SB, Saidha S, Newsome SD, Ratchford JN, Levy M, Ford E, Crainiceanu CM, Durbin MK, Oakley JD, Meyer SA, Frohman EM, Calabresi PA.  Brain. 2012 Feb;135(Pt 2):521-33. Epub 2011 Oct 17.

Microcystic macular oedema, thickness of the inner nuclear layer of the retina, and disease characteristics in multiple sclerosis: a retrospective study, Dr Shiv Saidha MRCPI, Elias S Sotirchos MD, Mohamed A Ibrahim MD, Ciprian M Crainiceanu PhD, Jeffrey M Gelfand MD, Yasir J Sepah MBBS, John N Ratchford MD , Jiwon Oh FRCPC, Michaela A Seigo ScB, Scott D Newsome DO, Prof Laura J Balcer MD, Prof Elliot M Frohman MD, Ari J Green, Quan D Nguyen MD, Prof Peter A Calabresi MD.  The Lancet Neurology, Early Online Publication, 4 October 2012, doi:10.1016/S1474-4422(12)70213-2

Wide 3-dimensional macular ganglion cell complex imaging with spectral-domain optical coherence tomography in glaucoma. Morooka S, Hangai M, Nukada M, Nakano N, Takayama K, Kimura Y, Akagi T, Ikeda HO, Nonaka A, Yoshimura N.  Invest Ophthalmol Vis Sci. 2012 Jul 20;53(8):4805-12. Print 2012 Jul.

Additional references can be found at:

About Voxeleron

Voxeleron is a Silicon Valley-based software house specializing in the development and licensing of computer vision and machine learning software for industrial and medical applications.

Dr. Peter Calabresi and Dr. Robert Chang have no financial interest in Voxeleron LLC.