Artifacts in automatic retinal segmentation using different optical coherence tomography instruments
G. B. Bietti Eye Foundation, IRCCS, Rome, Italy. Retina (Philadelphia, Pa.)
(Impact Factor: 3.24).
04/2010; 30(4):607-16. DOI: 10.1097/IAE.0b013e3181c2e09d
The purpose of this study was to compare and evaluate artifact errors in automatic inner and outer retinal boundary detection produced by different time-domain and spectral-domain optical coherence tomography (OCT) instruments.
Normal and pathologic eyes were imaged by six different OCT devices. For each instrument, standard analysis protocols were used for macular thickness evaluation. Error frequencies, defined as the percentage of examinations affected by at least one error in retinal segmentation (EF-exam) and the percentage of total errors per total B-scans, were assessed for each instrument. In addition, inner versus outer retinal boundary delimitation and central (1,000 microm) versus noncentral location of errors were studied.
The study population of the EF-exam for all instruments was 25.8%. The EF-exam of normal eyes was 6.9%, whereas in all pathologic eyes, it was 32.7% (P < 0.0001). The EF-exam was highest in eyes with macular holes, 83.3%, followed by epiretinal membrane with cystoid macular edema, 66.6%, and neovascular age-related macular degeneration, 50.3%. The different OCT instruments produced different EF-exam values (P < 0.0001). The Zeiss Stratus produced the highest percentage of total errors per total B-scans compared with the other OCT systems, and this was statistically significant for all devices (P < or = 0.005) except the Optovue RTvue-100 (P = 0.165).
Spectral-domain OCT instruments reduce, but do not eliminate, errors in retinal segmentation. Moreover, accurate segmentation is lower in pathologic eyes compared with normal eyes for all instruments. The important differences in EF among the instruments studied are probably attributable to analysis algorithms used to set retinal inner and outer boundaries. Manual adjustments of retinal segmentations could reduce errors, but it will be important to evaluate interoperator variability.
Available from: Peter Maloca
- "This very high rate of CSI segmentation failure may also be due to the fact that the algorithm was primarily designed for linear propagating tissue, which may not be applicable to outer choroid or lamina fusca. Because many artifacts may not be easily visible on the final 3D- mapping of the choroid, it is absolutely recommended that the individual cross-sectional images of artifacts be controlled by the techniques mentioned in previous reports[23,24]. Furthermore , an automatic OCT data classifier, which monitors segmentation abnormalities, should be implemented in every OCT system. This classifier takes into account the signal-tonoise ratio and automatically prejudges images with respect to completeness (anterior, posterior, and lateral truncation) and blink error, which reduces the plethora of possible images. "
[Show abstract] [Hide abstract]
Our purpose was to compare the tumor sizes of small choroidal nevi using ultra-widefield imaging (UWF) and different optical coherence tomography systems.
Thirteen choroidal nevi were measured using automatic and manual segmentation techniques, including enhanced depth imaging spectral-domain optical coherence tomography (EDI-SDOCT) and 1050 nm swept source OCT (SSOCT), to compare to measurements obtained using the Optos projection ultra-widefield fundus (UWF) imaging technique. Segmentation artifacts were evaluated for all 13 cases, alongside an additional 12 choroidal nevi, using SSOCT.
In tumor eyes, segmentation artifacts for the choroid-sclera interface were found in 42 % of SSOCT scans. EDI-SDOCT can underestimate tumor dimensions and differs up to -8.41 % compared to UWF imaging and by 1.25 % compared to SSOCT cases. The horizontal length of the nevi showed an average difference between EDI-SDOCT and SSOCT of ± 9.38 %. Measured markers showed an average difference in length of ± 12.51 %. The average tumor thickness showed a difference of ± 11.47 %. Comparisons between EDI-SDOCT/UWF, SSOCT/EDI-SDOCT, and marker EDI-SDOCT/SSOCT showed significant mean differences of -122 μm (CI: -212 to -31 μm, p = 0.013), 134 μm (CI: 65-203 μm, p = 0.0012), and -193 μm (CI: -345 to -41 μm, p = 0.017), whereas SSOCT/UWF showed no significant difference with a measurement of 13 μm (CI: -69-95 μm, p = 0.74).
Automatic segmentation of nevi requires much caution, because a choroidal tumor can trigger many artifacts. It would be beneficial to monitor choroidal nevi using the same type of OCT technology, because a tumor is displayed differently.
Available from: Bartosz L Sikorski
- "The OCT data is automatically segmented in order to generate the above maps (Figure 2). When interpreting these maps, one should bear in mind that the artefacts may occur during segmentation, which will lead to improper retinal thickness measurements [10, 11]. Artefacts may arise as a result of poor image quality, eye movement during measurements, and retinal pathologies interfering with automated segmentation (e.g., retinal pigment epithelial detachment, subretinal fluid, fibrosis, or haemorrhage). "
[Show abstract] [Hide abstract]
ABSTRACT: Diabetic maculopathy (DM) is one of the major causes of vision impairment in individuals with diabetes. The traditional approach to diagnosis of DM includes fundus ophthalmoscopy and fluorescein angiography. Although very useful clinically, these methods do not contribute much to the evaluation of retinal morphology and its thickness profile. That is why a new technique called optical coherence tomography (OCT) was utilized to perform cross-sectional imaging of the retina. It facilitates measuring the macular thickening, quantification of diabetic macular oedema, and detecting vitreoretinal traction. Thus, OCT may assist in patient selection with DM who can benefit from treatment, identify what treatment is indicated, guide its implementing, and allow precise monitoring of treatment response. It seems to be the technique of choice for the early detection of macular oedema and for the followup of DM.
- "The segmentation algorithm of each machine determined the boundaries. When measuring CMT, the Cirrus HD-OCT measured from the internal limiting membrane to the retinal pigment epithelium (RPE), while the Spectralis HRA + OCT measured from the internal limiting membrane to Bruch's membrane by distinguishing the RPE from Bruch's membrane. The frequency and severity of segmentation errors were higher in the Spectralis HRA + OCT than in the Cirrus HD-OCT, mainly because the RPE remains more or less intact compared to Bruch's membrane, which is invaded and destroyed by CNV. "
[Show abstract] [Hide abstract]
To evaluate frequency and severity of segmentation errors of two spectral-domain optical coherence tomography (SD-OCT) devices and error effect on central macular thickness (CMT) measurements.
Materials and Methods:
Twenty-seven eyes of 25 patients with neovascular age-related macular degeneration, examined using the Cirrus HD-OCT and Spectralis HRA + OCT, were retrospectively reviewed. Macular cube 512 × 128 and 5-line raster scans were performed with the Cirrus and 512 × 25 volume scans with the Spectralis. Frequency and severity of segmentation errors were compared between scans.
Segmentation error frequency was 47.4% (baseline), 40.7% (1 month), 40.7% (2 months), and 48.1% (6 months) for the Cirrus, and 59.3%, 62.2%, 57.8%, and 63.7%, respectively, for the Spectralis, differing significantly between devices at all examinations (P < 0.05), except at baseline. Average error score was 1.21 ± 1.65 (baseline), 0.79 ± 1.18 (1 month), 0.74 ± 1.12 (2 months), and 0.96 ± 1.11 (6 months) for the Cirrus, and 1.73 ± 1.50, 1.54 ± 1.35, 1.38 ± 1.40, and 1.49 ± 1.30, respectively, for the Spectralis, differing significantly at 1 month and 2 months (P < 0.02). Automated and manual CMT measurements by the Spectralis were larger than those by the Cirrus.
The Cirrus HD-OCT had a lower frequency and severity of segmentation error than the Spectralis HRA + OCT. SD-OCT error should be considered when evaluating retinal thickness.
Data provided are for informational purposes only. Although carefully collected, accuracy cannot be guaranteed. The impact factor represents a rough estimation of the journal's impact factor and does not reflect the actual current impact factor. Publisher conditions are provided by RoMEO. Differing provisions from the publisher's actual policy or licence agreement may be applicable.