Ocular fundus reference images from optical coherence tomography

Abstract

Two-dimensional images computed from three-dimensional optical coherence tomography (OCT) data are intrinsically aligned with it, allowing to accurately position a retinal OCT scan within the ocular fundus. In this work, we aim to compute an OCT fundus reference image with improved retinal vasculature extension and contrast over traditional approaches. Based on the shadow casted by hemoglobin on the outer layers of the retina, we compute three independent images from the OCT volumetric data (including the traditional fundus reference image). Combining these images, a fourth one is created that is able to outperform the other three, both quantitatively and qualitatively (as evaluated by retina specialists). The vascular network extension provided by this method was also compared with widely used fundus imaging modalities, showing that it is similar to that achieved with color fundus photography. In this way, the proposed method is an important starting point to the segmentation of the vascular tree and provides users with a detailed fundus reference image.

Introduction

Several pathologic conditions of the retina have been shown to be correlated to abnormal vascular patterns [1]. The retina is regarded as a window through which to reflect on the vascular system, and changes seen in the retinal microvasculature have been found to be related to several cardiac diseases [2], [3]. Hence, the detection and quantitative characterization of the human retina vascular network are important research areas. Properties such as arteriolar diameter at bifurcation and tortuosity have been associated with death from ischemic heart disease [4], and arteriolar narrowing has been associated with increased mortality from stroke [4]. Furthermore, several characteristics of the retinal vascular network have been shown to be connected to the state of the cerebral vascular system [5], [6]. Also, studies have reported consistent associations between vascular abnormalities and several disorders such as diabetes [7] and hypertension [8], [9].

Extensive work has been done in this field, mainly based on two widely used imaging modalities: color fundus photography (CFP) and fluorescein angiography (FA). To facilitate ocular fundus imaging and improve CFP image quality, eyedrops are used to dilate the pupil, making this imaging modality minimally invasive. On the other hand, sodium fluorescein (NaFl, a fluorescent dye) is administered intravenously before the imaging to highlight the vascular network, making FA invasive.

Optical coherence tomography (OCT) is an imaging modality capable of noninvasively imaging tissue microstructure in vivo and in situ. It is an important tool in the diagnosis of ocular pathologic conditions and has been used extensively in research and in clinical practice. This modality uses the interferometer principle and is extensively described in the literature [10], [11].

The introduction of the commercial high-definition spectral-domain OCT (SD-OCT) models has made it possible to obtain large, high-resolution 3D scans of the retina (Fig. 1) with short acquisition time. An accurate fundus reference image computed from OCT data is still necessary, even though several OCT systems use a fundus image from a scanning laser ophthalmoscope (SLO) as a surrogate reference. Although highly important, the SLO image is not necessarily aligned with the OCT volume, as SLO images are taken after the OCT scan and not simultaneously. In addition, the SLO does not convey the same information as the OCT data because of the different working principle. By adding up the 3D OCT data depth-wise, a 2D reference image can be computed. From a vascular network point of view, this image correlates to the fundus images obtained from CFP or FA. As opposed to SLO, these OCT fundus images are created from OCT data and, as such, are intrinsically aligned to the volume.

An accurate reference image is highly important, as shown by the several approaches that have been proposed to achieve the co-registration of OCT scans [12], [13], [14]. These approaches have been performed to co-register OCT scans taken at different points in time and to develop complementary imaging modalities (expected to lead to a deeper insight into the pathophysiology of retinal lesions) [12], [13], [14].

From the 2D fundus reference image of the OCT 3D data, one can also segment important retinal features that complement the 3D information, such as the fovea, the optic disk, and the retinal vascular network, similarly to CFP [14], [15], [16], [17]. On standard SD-OCTs, the retinal blood vessels are not directly visible. Instead, hemoglobin absorbs the light at the wavelength used by these systems. Consequently, the signal is attenuated at the structures below the vessels [10], [18], [19] (Fig. 2). This effect is well known and has been used to obtain the 2D vascular network from the 3D OCT data [20], [21].

In this work, we propose a method to generate a fundus reference image from 3D OCT data with improved vascular network extension and contrast. It combines data from three distinct 2D images computed from the 3D OCT data – the mean value fundus (MVF), the expected value fundus (EVF), and the error to local median (ELM) – to create the principal component fundus (PCF) reference image. To assess the extent of vascular network provided by the proposed OCT fundus reference image, we compared it with those provided by raw CFP and FA. This comparison is based on the visible vascular network as detected by a human grader who manually segmented the vascular network in the three imaging modalities. FA is the gold standard for ocular vascular network imaging and is used in this work to establish the total vascular network within the imaged region.