Contributions
Abstract: P559
Type: Poster
Abstract Category: Pathology and pathogenesis of MS - OCT
Background: Intra-retinal layer thickness measurements assessed by optical coherence tomography (OCT) are increasingly implemented as outcome parameters in clinical MS trials. Insufficient image quality might interfere with segmentation results and hamper detection of subtle intra-retinal changes in longitudinal clinical trials.
Objective: To compare segmentation quality of current retinal layer segmentation algorithms in relation to signal quality.
Methods: Macula volume scans of 12 healthy controls were acquired with Spectralis and Cirrus OCT devices applying different settings for focus and image averaging (Spectralis only) to simulate a broad variety of noise levels. A device-independent signal-to-noise ratio (SNR) parameter was established and used to classify scans as high (SNR >71dB), medium (71 - 46dB) and low (< 46dB) quality scans. Layer segmentation was performed with the device"s build-in algorithms (HEYEX and Cirrus software) and device-independent software (Iowa Reference Algorithm (IRA) and AUtomated Retinal Analysis (AURA) tools). Segmentation results were classified by an experienced rater and layer thicknesses were compared to scans with the highest quality. Correlation and repeatability of the algorithms were investigated in dependency of SNR levels.
Results: On high and medium SNR scans, segmentation results for all algorithms were similarly good with only few exceptions, but low SNR images caused severe errors in all segmentation algorithms. The algorithms generally showed good repeatability in high and medium SNR scans for all layers (Intraclass correlation coefficient (ICC): 0.82 - 0.99) with exception of the outer plexiform layer (IRA ICC: 0.62, HEYEX ICC: 0.71). The ganglion cell layer, inner plexiform layer and the complex of both showed stable segmentation results between scans with high and medium SNR (mean deviation (MD) to best scan =3µm), while low SNR resulted in higher MD (=10µm). The smallest differences were detected for the inner nuclear layer (MD=2µm) with the exception of low SNR scans segmented with HEYEX (MD=8µm). Correlations between algorithms and devices were high and gradually decreased with lower SNR.
Conclusions: The quality of OCT scans and the choice of suitable segmentation algorithms are crucial for generating reliable results in clinical MS trials. The presented device-independent SNR parameter and the analysis of current segmentation algorithms provide guidance to develop appropriate study designs.
Disclosure: Timm Oberwahrenbrock is supported by EXIST-Forschungstransfer (German Federal Ministry for Economic Affairs and Energy, 03EFEBE079) and has received speaker honorary from TEVA, Germany and Bayer, Germany.
Rebecca Jost has nothing to disclose.
Hanna Zimmermann has received speaker honorary from TEVA, Germany and Bayer, Germany
Ingeborg Beckers has nothing to disclose.
Friedemann Paul is supported by the Deutsche Forschungsgemeinschaft (DFG EXC257), the Bundesministerium für Bildung und Forschung (BMBF Competence Network Multiple Sclerosis) and the Guthy Jackson Charitable Foundation, he has received travel grants, research support and personal compensation for activities with Alexion, Bayer, MerckSerono, Teva, Sanofi Genzyme, MedImmune, Chugai, BiogenIdec and Novartis.
Alexander U. Brandt received funding for research from Novartis, Biogen, BMWi, BMBF and consulting fees unrelated to this study from Biogen, Novartis, Teva, Nexus, and Motognosis.
Abstract: P559
Type: Poster
Abstract Category: Pathology and pathogenesis of MS - OCT
Background: Intra-retinal layer thickness measurements assessed by optical coherence tomography (OCT) are increasingly implemented as outcome parameters in clinical MS trials. Insufficient image quality might interfere with segmentation results and hamper detection of subtle intra-retinal changes in longitudinal clinical trials.
Objective: To compare segmentation quality of current retinal layer segmentation algorithms in relation to signal quality.
Methods: Macula volume scans of 12 healthy controls were acquired with Spectralis and Cirrus OCT devices applying different settings for focus and image averaging (Spectralis only) to simulate a broad variety of noise levels. A device-independent signal-to-noise ratio (SNR) parameter was established and used to classify scans as high (SNR >71dB), medium (71 - 46dB) and low (< 46dB) quality scans. Layer segmentation was performed with the device"s build-in algorithms (HEYEX and Cirrus software) and device-independent software (Iowa Reference Algorithm (IRA) and AUtomated Retinal Analysis (AURA) tools). Segmentation results were classified by an experienced rater and layer thicknesses were compared to scans with the highest quality. Correlation and repeatability of the algorithms were investigated in dependency of SNR levels.
Results: On high and medium SNR scans, segmentation results for all algorithms were similarly good with only few exceptions, but low SNR images caused severe errors in all segmentation algorithms. The algorithms generally showed good repeatability in high and medium SNR scans for all layers (Intraclass correlation coefficient (ICC): 0.82 - 0.99) with exception of the outer plexiform layer (IRA ICC: 0.62, HEYEX ICC: 0.71). The ganglion cell layer, inner plexiform layer and the complex of both showed stable segmentation results between scans with high and medium SNR (mean deviation (MD) to best scan =3µm), while low SNR resulted in higher MD (=10µm). The smallest differences were detected for the inner nuclear layer (MD=2µm) with the exception of low SNR scans segmented with HEYEX (MD=8µm). Correlations between algorithms and devices were high and gradually decreased with lower SNR.
Conclusions: The quality of OCT scans and the choice of suitable segmentation algorithms are crucial for generating reliable results in clinical MS trials. The presented device-independent SNR parameter and the analysis of current segmentation algorithms provide guidance to develop appropriate study designs.
Disclosure: Timm Oberwahrenbrock is supported by EXIST-Forschungstransfer (German Federal Ministry for Economic Affairs and Energy, 03EFEBE079) and has received speaker honorary from TEVA, Germany and Bayer, Germany.
Rebecca Jost has nothing to disclose.
Hanna Zimmermann has received speaker honorary from TEVA, Germany and Bayer, Germany
Ingeborg Beckers has nothing to disclose.
Friedemann Paul is supported by the Deutsche Forschungsgemeinschaft (DFG EXC257), the Bundesministerium für Bildung und Forschung (BMBF Competence Network Multiple Sclerosis) and the Guthy Jackson Charitable Foundation, he has received travel grants, research support and personal compensation for activities with Alexion, Bayer, MerckSerono, Teva, Sanofi Genzyme, MedImmune, Chugai, BiogenIdec and Novartis.
Alexander U. Brandt received funding for research from Novartis, Biogen, BMWi, BMBF and consulting fees unrelated to this study from Biogen, Novartis, Teva, Nexus, and Motognosis.