Blood Vessel Tree Reconstruction in Retinal OCT Data Gazárek J, Kolář R, Jan J, Odstrčilík J, Taševský P Department of Biomedical Engineering, FEEC, Brno University of Technology xgazar03@stud.feec.vutbr.cz Initial study of blood vessels reconstruction in the retinal optical coherence tomography (OCT) data is reported in this paper. Reconstruction and segmentation of the blood vessel tree is an important task for image registration, because it carries spatial information about the OCT B-scans positions. The dark shadows created by blood-vessels in the retinal pigment epithelium (RPE) shown in the OCT scans offer a possibility for blood vessel tree reconstruction. The main part of this paper is therefore focused on RPE layer detection and alignment. Promising results have been obtained for OCT data concerning several patients in this initial study. 1 Introduction The optical coherence tomography (OCT) has found many applications in medical and biological sciences, particularly in clinical ophthalmology. One of the most important applications is the early glaucoma diagnosis based either on the examination of the retinal nerve fibre layer from cross-sectional scans [1] or based on the segmentation of optic disc cup and rim [2]. Another important clinical application is OCT-based diagnosis of the age related macular degeneration [3]. Nowadays, the OCT devices work in the spectral domain, which enables faster acquisition of particular cross-sectional tomographic images (B-scans) of retinal structures [4]. A data set from a single measurement is therefore three dimensional (3D). Although the data acquisition is fast, the patient and eye movement can cause artefacts, which can disturb the workflow of the following data processing and analysis. This paper is focused on the retinal blood vessel tree reconstruction and segmentation in the OCT data [5]. The blood vessels create important features for monomodal or multimodal registration of retinal images; for example OCT data and images taken by the fundus camera [6] or scanning laser ophthalmoscopy (SLO) images [7]. 2 Methods The positions of the blood vessels can be recognized in each B-scan as dark stripes, which are caused by shadows generated by blood vessels in outer part of retina. This property can be used for detection. The highest contrast between the blood vessels shadows and the background can be observed in the retinal pigment epithelium (RPE) and surrounding bright layers. Therefore, the main segmentation task is the RPE detection. The proposed approach for RPE segmentation is depicted in Fig. 1. Fig. 1. The flowchart of proposed method for RPE detection 287
The typical B-scan without significant distortion with several marked structures is shown in Fig. 2. The retinal pigment epithelium (RPE) is an intraretinal layer with high optical reflectivity, which can be relatively easily detected in a case of low axial warping. Fig. 2. One B-scan with several retinal structures important for presented application. ILM: internal limiting membrane, RNFL: retinal nerve fibre layer, RPE: retinal pigment epithelium The proposed approach was tested on 8 volume sets from 4 eyes. The volume B-scans had been obtained by OCT (Heidelberg Engineering Spectralis) and has resolution 496 768 pixels. They were acquired with high density B-mode scanning (from 61 to 169 B-scans per field of view). Fig. 3 shows two distorted B-scans with axially warped retinal layers. This kind of distortion is caused by the high frequency eye movement, which is impossible to track and eliminate [8]. Fig. 3. One B-scan with low level of spatial warping; An example of B-scan with high spatial distortion. 2.1 Preprocessing The preprocessing step consists of median filtering for speckle noise elimination. The mask size has been determined experimentally to obtain sharp edges: 21x21 pixels. This filter has been applied two times to obtain more noise-free images. The filtered image was consequently filtered with convolution kernel approximating the gradient in vertical direction. 288
Fig. 4. The gradient images corresponding to images in Fig. 3. 2.2 RPE localization It has been observed that the strongest edge in each column arises from the internal limiting membrane (ILM) or RPE layer for majority of A-scans. Therefore two strongest reflections can be easily detected taking the maximum intensity for each A-scan in the gradient image. The second strongest edge is considered if the gradient is above detected threshold (0.01), which helps to eliminate detection of the weak edges inside the optic disc. The next step consists of moving averaging of the spatial position of detected RPE edge. The size of the averaging window has been set to 5 pixels. The next step helps to eliminate those edge pixels or edge fragments, which don t belong to RPE layer. A priori information about the continuity of RPE is used defining the local tangent (LT) as: LTk = positionx k +1 positionx k, k = 1,2 K n 1 positiony k +1 positiony k, where n is a number of connected edge points (fragments), positionxk+1 (positionyk+1) is X (Y) coordinate of the nearest fragment to the neighbouring fragment pixel kpositionx (kpositiony). If LT value is higher than 1.5, the shorter fragment is erased. This algorithm works iteratively until the LT values are lower than 1.5. Two results of this step are shown in Fig. 5. Fig. 5. Two examples of RPE localization. The presence of optic disc ( complicates the RPE detection. 2.3 Postprocessing After removing the outlying detected points, the cubic interpolation is used to compute the missing points in RPE layer. For smoothing the polynomial approximation of 20th order is used. This higher order is necessary for successful tracking of high spatial RPE variation in distorted images. The results are discussed in the next section. 289
3 Results An example of two results is presented in Fig. 6a,b. Successfully detected spatial variation can be seen, which corresponds to an approximate position of the RPE layer. Nevertheless, the precise detection of the RPE is not the main task of this work. The main reason for the RPE detection is the reconstruction and segmentation of the blood vessels, which is simpler for the aligned B-scans (see Fig. 6c,d). c) d) Fig. 6. a, Examples of two B-scans and detected shape of spatial warping; c, d) corresponding images after warping elimination. B-scan on c) presents an example of blood vessel detection. The reconstructed blood vessel tree local luminosity can be obtained by column-wise averaging of 10 pixels belonging to RPE layer. This operation eliminates the noise and in the end increases the blood vessel contrast in the reconstructed image. The results, i.e. blood-vessel images are presented here for two different eyes. Fig. 7 shows the successfully reconstructed tree without any artefacts and the corresponding colour fundus image. Fig. 8 shows the results with several uncorrected moving artefacts, which are marked by red arrows. These artefacts are caused by false RPE detection due to several effects, as: - higher level of noise in case that all B-scans do not have the same signal-to-noise ratio, which is probably caused by different number of individual B-scans used for averaging [4]; - absence of RPE - in the region of optic disc or macula; - detection of false edges corresponding to different retinal layers with strong reflection; - wrong B-scan acquisition the spatial warping can be so high, that the retinal layers lie outside of the acquired image depth range, in some cases; 290
Fig.7. Reconstructed blood-vessel OCT image and the corresponding colour fundus image Fig.8. Another example of reconstructed blood-vessel image with the corresponding colour fundus image 291
4 Conclusions A new method for blood vessel tree reconstruction based on 2D detection of RPE in OCT B-scans was proposed. The algorithm was tested on 8 cases of eye volume scans with 783 B-scans in total. The detected RPE layer has been subjectively evaluated and classified into 4 classes: 1. correct RPE detection: 92.46% (724 B-scans) 2. small distortion of RPE detection: 3,32% (26 B-scans) 3. high distortion of RPE detection: 1,92% (15 B-scans) 4. low-quality B-scans 2,30% (18 B-scans) The obtained 2D image of the vessel tree enables registering it with the corresponding fundus-camera retinal image, and consequently it is possible to find a position of each OCT B-scan in the fundus-camera image. This way, the objective quantitative evaluation of the retinal neural layer (RNL) obtained by the OCT method may be compared with the qualitative conclusion on the local presence of the RNL provided by the texture analysis based detection in the fundus-camera image. A future development of this method will cover 3D RPE segmentation, which will utilize also the correlation along the 3 rd (transversal) dimension. Acknowledgement This work has been supported by the national research center DAR (Data, Algorithms and Decision making) project no. 1M0572 coordinated by the Institute of Information Theory and Automation, Academy of Science, Czech Rep. and partly also by the institutional research frame no. MSM 0021630513; both grants sponsored by the Ministry of Education of the Czech Republic. The authors highly acknowledge the cooperation with University Eye Hospital Erlangen-Nurnberg (R.Tornow, R. Laemmer, Ch. Mardin), through which also the test set of images was provided. References [1] Trip S.A, Schlottmann P.G, Jones S.J et al. Retinal nerve fibber layer axonal loss and visual dysfunction in optic neuritis. American Journal of Ophthalmology, Vol. 140, No. 6, 2005. [2] Lee K, Niemeijer M, Garvin M.K et al. Segmentation of the optic disc in 3-d oct scans of the optic nerve head. Medici Imaging, IEEE Transactions on, Vol. 29, No. 1, pp. 159-168, Jan 2010. [3] Coscas G, Coscas F, Zourdani A et al. Optical coherence tomography and armd. J Fr Ophtalmol., Vol. 27, pp. 3S7 3S30, 2004. [4] Drexler W, Fujimoto J.G. Optical Coherence Tomography. Springer, 2008. [5] Fercher A.F, Drexler W, Hitzenberger C.K et al. Optical coherence tomography - principles and applications Reports on Progress in Physics, Vol.66, pp. 239-303, 2003. [6] Kolar R, Tasevsky P. Registration of 3D Retinal Optical Coherence Tomography Data and 2D Fundus Images. Accepted to Workshop on Biomedical Image Registration 2010, Luebeck, Germany. [7] Ricco S, Chen M, Ishikawa H et al., Correcting motion artifacts in retinal spectral domain optical coherence tomography via image registration. Medical Image Computing and Computer-Assisted Intervention MICCAI 2009, pp. 100-107, October 2009. [Online]. Available: http://dx.doi.org/10.1007/978-3-642-04268-3_13. [8] Fuller A.R, Zawadzki R.J, Choi S et al., Segmentation of three-dimensional retinal image data. IEEE Transactions on Visualization and Computer Graphics, Vol. 13, No. 6, pp. 1719-1726, November 2007. 292