Purchase this article with an account.
Kyungmoo Lee, Gabrielle HS Buitendijk, Hrvoje Bogunovic, Henriet Springelkamp, Albert Hofman, Andreas Wahle, Milan Sonka, Johannes R Vingerling, Caroline C W Klaver, Michael David Abramoff; Validation of Segmentability Index for Automated Prediction of Segmentation Reliability in SD-OCT Scans. Invest. Ophthalmol. Vis. Sci. 2015;56(7 ):5292.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Though some commercially available SD-OCT devices provide an image quality index, these are not available on all, nor are they consistent across all SD-OCT devices. More importantly, they do not predict the reliability of different structures, such as retinal layers, in the SD-OCT scan volume. We have developed a new metric, Segmentability Index (SI), to predict the reliability of the segmentation.
3808 macular SD-OCT (3D OCT-1000 , Topcon Europe, Netherlands) volumes (512 × 128 × 650 voxels, 6.0 × 6.0 × 2.3 mm3) were obtained from both eyes of 1128 subjects (74.7 ± 8.3 years, 41% male) randomly selected from the population-based Rotterdam study. A subset consisting of 50 OCT volumes with successful RNFL segmentations and 50 volumes with failed segmentations (Iowa Reference Algorithms: https://www.iibi.uiowa.edu/content/iowa-reference-algorithms-human-and-murine-oct-retinal-layer-analysis-and-display) was created. The SI is obtained from a random forest regressor based on 12 features: aggregate on-surface costs of inner/outer RNFL and outer RPE surfaces, mean and standard deviations of whole OCT voxel intensities, and whole edge-based costs of dark-to-bright and bright-to-dark transitions. A 10-fold cross-validation on these 100 volumes was compared to those obtained from a random forest regressor trained by a well-known quality index, maximum tissue contrast index (mTCI) based on the OCT voxel intensity distribution , using receiver operating characteristic (ROC) analysis.<br />  Huang Y, Gangaputra S, Lee KE, et al. Signal quality assessment of retinal optical coherence tomography images. Invest Ophthalmol Vis Sci. 2012;53(4):2133-41.
Areas under the curves (AUCs) of the mTCI and SI [95% confidence intervals (CIs)] are 0.722 [0.623, 0.822] and 0.892 [0.828, 0.957], respectively (Fig. 1), so that the SI ROC curve is significantly better than the mTCI curve (p < 0.001).
The SI is well suited to identify OCT scans that provide successful automated RNFL segmentations and those that do not allow reliable segmentations. Using the SI, scans with unreliable segmentations can be identified allowing more reliable automated segmentation analyses in population studies, in a device-independent manner.
This PDF is available to Subscribers Only