Automatic and manual segmentation of healthy retinas using high-definition optical coherence tomography

Transcription

1 Automatic and manual segmentation of healthy retinas using high-definition optical coherence tomography Isabelle Golbaz, 1 Christian Ahlers, 1 Nina Goesseringer, 2 Geraldine Stock, 1 Wolfgang Geitzenauer, 1 Christian Pru nte 1 and Ursula Margarethe Schmidt-Erfurth 1 1 Department of Ophthalmology, Medical University of Vienna, Vienna, Austria 2 Department of Plastic and Reconstructive Surgery, Rudolfstiftung Hospital, Vienna, Austria ABSTRACT. Purpose: This study compared automatic- and manual segmentation modalities in the retina of healthy eyes using high-definition optical coherence tomography (HD-OCT). Methods: Twenty retinas in 20 healthy individuals were examined using an HD-OCT system (Carl Zeiss Meditec, Inc.). Three-dimensional imaging was performed with an axial resolution of 6 lm at a maximum scanning speed of A-scans second. Volumes of mm were scanned. Scans were analysed using a matlab-based algorithm and a manual segmentation software system (3D-Doctor). The volume values calculated by the two methods were compared. Results: Statistical analysis revealed a high correlation between automatic and manual modes of segmentation. The automatic mode of measuring retinal volume and the corresponding three-dimensional images provided similar results to the manual segmentation procedure. Both methods were able to visualize retinal and subretinal features accurately. Conclusions: This study compared two methods of assessing retinal volume using HD-OCT scans in healthy retinas. Both methods were able to provide realistic volumetric data when applied to raster scan sets. Manual segmentation methods represent an adequate tool with which to control automated processes and to identify clinically relevant structures, whereas automatic procedures will be needed to obtain data in larger patient populations. Key words: 3-D rendering high-definition optical coherence tomography (HD-OCT) imaging manual and automatic segmentation OCT Acta Ophthalmol. 2011: 89: ª 2009 The Authors Journal compilation ª 2009 Acta Ophthalmol doi: /j x Introduction Optical coherence tomography (OCT) has become increasingly important in ophthalmology. Since the early 1990s, OCT technology has been subject to several technical improvements and has been successfully established in most retina centres as a result of its non-invasive nature (van Velthoven et al. 2007). A main advantage of OCT is its capability to display and localize discrete morphological changes in detail (Drexler & Fujimoto 2007; Leitgeb 2007). Thus, OCT has become a standard method in diagnostics of chorioretinal pathologies and its use now surpasses that of angiography. As a result of its considerable improvements in transversal resolution, Stratus OCT (Carl Zeiss Meditec, Inc., Dublin, California, USA) has become established in clinical practice. This device is able to scan the macula in a radial pattern, consistent with taking six single scans, and thus provides information on retinal thickness throughout the macular area. The resulting maps, in large part, reflect virtual data because the spaces between the six scans must be interpolated by an algorithm. However, conventional OCT systems are limited by their low scanning speed and low axial resolution. In addition, mistakes in the automatic detection of retinal thickness have compromised the 185

2 data provided by conventional OCT systems (Sadda et al. 2006). The introduction of fast raster scanning high-definition OCT (HD-OCT) into the clinic has improved retinal imaging because it allows for a significant increase in scanning speed. This advantage over time domain systems enables HD-OCT to visualize all relevant locations within the macular area realistically because minimizing the time needed to assess datasets is crucial to avoid artificial results provoked by eye movement or other disturbing influences. The increased resolution enables a more exact identification of retinal pathologies (Schmidt-Erfurth et al. 2005). Furthermore, the higher resolution allows for better discrimination between individual retinal layers (Wojtkowski et al. 2005). Admittedly, the methods used for retinal thickness analyses in current HD-OCT systems are still based on algorithms, which automatically identify the interface between distinct retinal structures, and in which errors may occur during the segmentation process. Such failures have shown, firstly, that the parameter retinal thickness must be evaluated using methods which are likely to avoid algorithm errors and, secondly, normative data with which upcoming generations of algorithms can be compared must be established. Furthermore, it is clearly useful to establish a measurement system that does not depend on a specific scanning device. In this clinical study, manual segmentation modes were compared with automated procedures in healthy retinas in order to obtain an independent, FDA-approved methodology and to compare the results of manual and automatic procedures. Materials and Methods Twenty normal eyes in 20 healthy volunteers were included. Any retinal or ophthalmologic disease in the candidate s past medical history led to exclusion. The subjects average age was 29 years (range years). All patients were White. The group consisted of eight men and 12 women. The macular region was examined and analysed using HD-OCT. Best corrected visual acuity (BCVA) was measured with ETDRS (Early Treatment Diabetic Retinopathy Study) charts. All eyes were required to achieve Snellen BCVA of Refractory aberrations were not to exceed ± 1 D. Subjects underwent an extensive informed consent procedure and were asked for written consent. The protocol of the study was approved by the local ethics committee and adhered to the standards of the Declaration of Helsinki. The examinations were performed at the Department of Ophthalmology, Medical University of Vienna, Austria. High-definition optical coherence tomography A prototype of the Cirrus HD-OCT (Carl Zeiss Meditec, Inc.) was used to provide three-dimensional (3-D) imaging in 20 healthy eyes within the macular region. The device is a second-generation frequency domain HD-OCT system with an axial resolution of 6 lm and a maximum scanning speed of A-scans second. The scanning area measured 6 6 mm. A macular raster scan was selected for quantitative analysis. Automatic segmentation Automatic segmentation and volumetric analysis were performed using matlab-based software (MathWorks, Inc., Natick, Massachusetts, USA) and algorithms which were designed to automatically delineate the internal limiting membrane (ILM) and the retinal pigment epithelium (RPE) layer. This software was developed for Cirrus HD-OCT and has been used as a routine tool. Retinal thickness and volume are calculated by measuring the distance between the ILM and the contour of the RPE layer. The algorithm finds the ILM by thresholding the image at an adaptively defined level that separates the retina from vitreous material. The retinal image below the ILM location is then smoothed and the site of the most highly reflected intensity identified. The strongest edge below this brightest focus is detected and the resulting surface is laterally smoothed to give a result for the RPE (Fig. 1). Manual segmentation Manual segmentation and volumetric analysis were performed using an imaging software (3D-Doctor Ò ; Able Software Corp., Lexington, Massaschusetts, USA) cited frequently in scientific publications and approved by the US Food and Drug Administration (FDA) for medical imaging and 3-D visualization (Ahlers et al. 2008). To apply manual segmentation, the set of 128 B-scans contained in a complete macular raster scan was transformed from raw data into bitmap format. These bitmaps were imported into the manual imaging software. Certified graders from the Vienna Reading Centre (VRC; Department of Ophthalmology, Medical University Inclusion criteria (B) (C) (D) Fig. 1. Automatic segmentation. Retinal thickness and volume were calculated by measuring the distance between the inner limiting membrane (ILM) (green line) and the contour of the retinal pigment epithelium (RPE) layer (red line). Correct detection of the retinal boundaries is therefore crucial to the precise measurement of retinal volume. (B) Three-dimensional image of the retinal surface (ILM) and the RPE. (C) Retinal thickness map. (D) Topographic map of retinal thickness. N 186

3 of Vienna) delineated the ILM and the contour of the RPE layer to obtain the retinal volume in each B-scan. Accurate identification of the RPE layer was ensured by marking the surface of the outermost hyperreflective band within the outer photoreceptor and RPE complex (Fig. 2A). Each grader underwent a certification process in OCT grading at the VRC before being approved to grade images for this study. Manual segmentation was performed for a total of 2560 B-scans because 128 single B-scans represented one macular raster scan. Manual segmentation of one complete raster scan took the examiner approximately 2 3 hours. Three-dimensional images of retinal and macular architecture, as well as of retinal thickness, were reconstructed based on the complete set of segmented B-scans (Fig. 2). Volume calculation and statistics In order to compare the reliability of the two different methods of analysing retinal volume, automatically computed measurements were compared with measurements performed by manual segmentation. (B) Scatterplot analyses were used to evaluate the performance of each method, in addition to correlation and comparative analysis. A Bland Altman plot (Analyse-it Ò Standard Edition; Analyse-it Software Ltd, Leeds, UK) was used to visualize and describe the agreement between the two measurement methods. A p-value of < 0.05 was considered to be statistically significant. spss for Windows Version 15.0 (SPSS, Inc., Chicago, Illinois, USA) was used for the statistical analysis. Results Two segmentation methods were compared using HD-OCT scans of healthy retinas. Scans were analysed by an automated matlab-based software and an FDA-approved manual segmentation software previously used in radiology. The 3-D illustration of the retinal surface showed the physiological appearance of the macular region in all cases. 3-D visualization of segmented structures Both methods were capable of visualizing segmentation results in a 3-D mode. Topographic images of the retinal thickness and the RPE layer could Fig. 2. Manual segmentation. The inner limiting membrane and the contour of the retinal pigment epithelium layer were identified and delineated manually (orange lines) in all 128 single images of the mm scanned area. Serial sequences of slices and the appropriate boundaries can be displayed as known from computed tomography or magnetic resonance imaging examinations. (B) Surface reconstruction of the previously segmented images allows for threedimensional analysis of the physiological retina. The foveal contour and the concave pattern are clearly visible. be calculated from a raster scan in automatic segmentation. As only physiological conditions were examined in this study, none of the subjects showed a significant elevation or alteration in the RPE pathology map. Manual segmentation was able to present a 3-D reconstruction which could be freely moved or rotated. Axial movement of different structures and camera angle could be determined individually. Animations could be shown and saved in either an AVI or QuickTime video format. Correct placement of the boundaries could be controlled in each B-scan using a flythrough video function, which consecutively played the B-scans and their individual markers. Results of volumetric analysis Correct placement of the boundaries between different layers could be controlled by a video technique as well as in the single B-scans in both methods. In this study, volume data for the macular region were quantified in cubic millimetres by the automatic method and in voxels by the manual segmentation method. A voxel is a volume element representing a value on a regular grid in 3-D space. This is analogous to a pixel, which is used to represent 2-D image data. Voxels are frequently used in the visualization and analysis of medical and scientific data. In order to better compare the two methods, a conversion factor was applied to the voxel values which allowed the results to be compared in cubic millimetre format. Mean retinal volume values reached mm 3 (standard deviation [SD] 0.43) in manual segmentation and mm 3 (SD 0.34) in automatic segmentation. (Figs 3 and 4). Scatterplot analysis revealed a significant correlation (p < , slope 0.99, intercept 0.27) between automatic and manual segmentation using 3-D Doctor. Bland Altman plot analysis showed mean differences close to zero with narrow confidence intervals (CIs) (95% CI ) 0.65 to 0.39), indicating a good agreement between the two segmentation methods. The automatic measurements of retinal volume and the corresponding 3-D images provided results similar to those of manual segmentation following adequate conversion. 187

4 Automatic segmentation Conclusions Identity This study was conducted in order to compare two methods of assessing retinal volume using HD-OCT scans of the central retina in healthy eyes. Scatter plot Manual segmentation Fig. 3. Scatterplot analysis of manual and automatic segmentation procedures. Difference (Manual segmentation - Automatic segmentation) Difference plot Identity Bias ( 0.13) 95% Limits of agreement ( 0.65 to 0.39) Mean of all Fig. 4. Bland Altman plot demonstrating the agreement between two segmentation methods Quantitative comparisons were carried out and topographic features were described. Certified OCT graders performed manual segmentation of the macular region using FDAapproved manual imaging software and matlab-based automatic segmentation routines, which have been previously described in the literature (Ahlers et al. 2007). The study showed that both methods are able to provide reproducible and comparable volumetric data when applied to retinal raster scans. A statistically significant correlation was proven between the manual analysis and the automatic segmentation results. In both methods, retinal volume data were obtained by measuring retinal thickness from the ILM surface down to the RPE layer. Both methods were able to provide topographic 3-D animations of the measured structures, as well as of the surface of the RPE. Compared with manual segmentation, automatic segmentation may be problematic in selected cases because the identification of layers at the level of the RPE sometimes fails to consistently detect the identical bands of the three hyperreflective segments visible within the photoreceptor RPE interface. Similar findings have been described for conventional OCT systems (Costa et al. 2004) (Fig. 5). As shown in this study, changes in retinal reflectivity can be analysed successfully by manual segmentation. Such analyses have been conducted in radiology and are now presented in ophthalmology for the first time. The identification and differentiation of pathophysiologically relevant layers or structures is important to the identification of realistic correlations between functional and morphologic parameters (Srinivasan et al. 2006a, 2006b; Ruggeri et al. 2007; Ahlers et al. 2008). Manual segmentation of intraretinal volumes using FDA-approved manual (B) Fig. 5. Automatic segmentation algorithm error. Algorithms may fail in the correct detection of the inner limiting membrane in automatic segmentation, as shown in this physiological retina. (B) False identification of the retinal layers at the level of the retinal pigment epithelium. 188

5 segmentation is a valuable tool, although the methodology is compromised by the huge amount of time needed to assess volumetric data composed of a high number of single scans. However, the method is able to reduce segmentation-based artefacts when operated by carefully trained personnel and can therefore serve to control novel automatic procedures before they are applied to larger patient populations in clinical trials. Moreover, manual segmentation might be used to obtain measurements in a limited number of patients in order to identify and characterize the most relevant factors in particular pathological processes and thereby help to develop automated processes that are capable of acquiring such data automatically. In conclusion, both the automatic as well as the manual segmentation solutions for HD-OCT raster scans provided reliable results when compared under physiological conditions. Despite the substantial effort required to perform manual segmentation, the method will be useful for controlling automated procedures and for identifying clinically relevant structures in the future. It will be particularly challenging for HD-OCT technology to analyse data from retinal morphologies which demonstrate different levels of pathology. In such cases, which are more relevant to clinicians, the composition of the neurosensory layer is altered and segmentation becomes difficult. Future comparison studies will identify the limits of automatic segmentation. References Ahlers C, Simader C, Geitzenauer W, Stock G, Stetson P, Dastmalchi S, Schmidt- Erfurth U (2007): Automatic segmentation in three-dimensional analysis of fibrovascular pigment epithelial detachment using high-resolution optical coherence tomography. Br J Ophthalmol 92: Ahlers C, Golbaz I, Stock G, Fous A, Kolar S, Pruente C, Schmidt-Erfurth U (2008): Time course of morphologic effects on different retinal compartments after ranibizumab therapy in age-related macular degeneration. Ophthalmology 115: Costa RA, Calucci D, Skaf M, Cardillo JA, Castro JC, Melo LA Jr, Martins MC, Kaiser PK (2004): Optical coherence tomography 3: automatic delineation of the outer neural retinal boundary and its influence on retinal thickness measurements. Invest Ophthalmol Vis Sci 45: Drexler W & Fujimoto JG (2007): Stateof-the-art retinal optical coherence tomography. Prog Retin Eye Res 27: Leitgeb RA (2007): Optical coherence tomography high resolution imaging of structure and function. Conf Proc IEEE Eng Med Biol Soc 1: Ruggeri M, Wehbe H, Jiao S, Gregori G, Jockovich ME, Hackam A, Duan Y, Puliafito CA (2007): In vivo three-dimensional high-resolution imaging of rodent retina with spectral-domain optical coherence tomography. Invest Ophthalmol Vis Sci 48: Sadda SR, Wu Z, Walsh AC, Richine L, Dougall J, Cortez R, LaBree LD (2006): Errors in retinal thickness measurements obtained by optical coherence tomography. Ophthalmology 113: Schmidt-Erfurth U, Leitgeb RA, Michels S et al. (2005): Three-dimensional ultrahighresolution optical coherence tomography of macular diseases. Invest Ophthalmol Vis Sci 46: Srinivasan VJ, Ko TH, Wojtkowski M et al. (2006a): Non-invasive volumetric imaging and morphometry of the rodent retina with high-speed, ultrahigh-resolution optical coherence tomography. Invest Ophthalmol Vis Sci 47: Srinivasan VJ, Wojtkowski M, Witkin AJ et al. (2006b): High-definition and threedimensional imaging of macular pathologies with high-speed ultrahigh-resolution optical coherence tomography. Ophthalmology 113: 2054.e1 14. van Velthoven ME, Faber DJ, Verbraak FD, van Leeuwen TG, de Smet MD (2007): Recent developments in optical coherence tomography for imaging the retina. Prog Retin Eye Res 26: Wojtkowski M, Srinivasan V, Fujimoto JG, Ko T, Schuman JS, Kowalczyk A, Duker JS (2005): Three-dimensional retinal imaging with high-speed ultrahigh-resolution optical coherence tomography. Ophthalmology 112: Received on October 22nd, Accepted on March 17th, Correspondence: Christian Pru nte MD Department of Ophthalmology Medical University of Vienna Waehringer Guertel Vienna 1090 Austria Tel: Fax: christian.pruente@meduniwien.ac.at 189