Keywords

1 Introduction

Retinal vascular morphology can represent an important biomarker for diseases like diabetes [1], hypertension [2] or arteriosclerosis [3], among others. In recent years, the introduction and popularization of Optical Coherence Tomography (OCT) as a non-invasive exploratory technique for the analysis of the eye fundus became a reality. This image modality allows the doctors to obtain images of the retinal tissues, including the presence and location of the retinal vascular structure [4]. These images enable the experts to make a clinical evaluation of the retinal vascular tree morphology, permitting the identification and analysis of different types of diseases.

In OCT imaging, the retinal vessels are visualized as structures that block the transmission of light and leave a shadow. In the literature, we can find many different approaches that solve this problem in classical retinographies. As reference, the method in [5] is based on the use of a Hessian multiscale enhancement filter to detect the vessels. [6] uses an approach based on Gaussian and Kalman filters. Other approach include the use of neural networks [7]. A similar aim was proposed in [8] where authors use a deformable contour model to identify the vasculature. However, few studies were proposed in OCT images, as the works in [9,10,11], but using as support the corresponding near-infrared reflectance retinography for the vascular detection process.

We propose a complete methodology for the automatic detection of retinal vessels using, only, the histological sections of the OCT images. The method segments the retinal layers and detects the vessel structures by the analysis of statistical features of the intensity profile obtained between these layers. Two approaches are considered for this purpose: one using a fixed threshold and a second using an adaptive threshold.

2 Methodology

The proposed methodology, represented in Fig. 1, is divided into four main steps: a first step, where the retinal layers are segmented using the input image; a second step, applying a preprocessing to enhance the characteristics of the vascular structures; a third step, where a set of statistical features is extracted; and a fourth step, where the vessels are finally detected. Each one of these steps is going to be discussed next.

Fig. 1.
figure 1

Main steps of the proposed methodology.

Full size image

2.1 Retinal Layer Segmentation

The proposed method receives, as input, an OCT image. This image corresponds to a histological section representing the morphology of the retinal layers. Using these images, the retinal layers are identified and segmented, delimiting the region where the vessels are placed. For this purpose, we follow the methodology proposed in [12], where the retinal layers are segmented using an active contour-based technique. In particular, two layers are considered in this work: the Pigment Epithelium Bruch’s Complex with the Choroid (RPE/C) and top boundaries of the Ellipsoid (M/E). Between these layers, a higher contrast is perceived in the areas where vascular shadows are present (Fig. 2).

Fig. 2.
figure 2

Input OCT image marked with the retinal layers considered in this work: (RPE/C) and (M/E).

Full size image

2.2 Image Pre-processing

Speckle noise is a common distortion that is frequently present in OCT images. For that reason we applied a pre-processing step, removing the noise and increasing the contrast of the vascular shadows in the region between the previously detected layers. Firstly, we apply a Gaussian Blur Filter to reduce noise in the image. Then, a top-hat transform is used vertically to increase the contrast of vascular shadows. Figure 3 illustrates the results obtained in this stage.

Fig. 3.
figure 3

Pre-processed image after noise removal and contrast enhancement.

Full size image

2.3 Statistical Feature Extraction

In this phase, we detect the vessel structures by the analysis of the intensity profiles between the layers of the retina. To achieve this, we calculate a signal where each point represents the mean \(\mu {_i}\) of the intensity \(I_j\) in each column \(c_i\) within the region of interest (ROI), as indicated in the Eq. 1, where n represents the number of elements in the column \(c_i\).

$$\begin{aligned} \mu {_i} = \frac{\sum _{j=0}^{n}I_j}{n} \end{aligned}$$
(1)

We calculate these values on the pre-processed image and within the ROI that is delimited by the layers (RPE/C) and (M/E). Figure 4 shows the graphic representing the mean intensity profile from the OCT images. Global mean \(\mu \) and standard deviation \(\sigma \) are also calculated using all the column means \(\mu _i\).

Fig. 4.
figure 4

Graphic representing the statistical features: \(\mu _i\), \(\mu \) and \(\sigma \).

Full size image

2.4 Vessel Detection

In this phase, we detect the vascular structures in the slices using the information of the column means \(\mu _i\). For that purpose, we use two strategies, fixed and adaptive threshold approaches, that will be detailed below.

Fixed Threshold. In this first approach, we consider that the values obtained for the means \(\mu {_i}\) of each column \(c_i\) approximate a Normal distribution. Therefore, we can consider that the atypical values (outlayers) of the left tail of the distribution identify the vascular shadows. Consequently, we elaborate a model that uses a fixed threshold \(th_{fix}\), as shown in Eq. 2, where \(Q_1\) represent the \(1^{st}\) quartile and \(Q_3\) and the represent the \(3^{rd}\) quartile. This value allows us to detect the outlayers that represent the vascular structures.

$$\begin{aligned} th_{fix} = \frac{3}{2}(Q_3 - Q_1) + Q_3 \end{aligned}$$
(2)

The results of this approach can be observed in Fig. 6(a), (c), where the values that are below the fixed threshold \(th_{fix}\) represent the vascular detections in the OCT image.

Adaptive Threshold. The fixed threshold approach presents, as main limitation, that the column means \(\mu _i\) do not always oscillate around the global mean \(\mu \). Instead of that, many times the column means tendency is deviated to bright and dark values due to local intensity alterations. These deformations can make that a global fixed threshold miss-detects vessels (when the tendency moves brighter) or produce many false positives (when the tendency moves darker). This motivated the second approach. Based on the principle that the vessel presence produces a significant intensity depression, this second approach uses the mathematical concepts of local maxima (peak) and minima (valley) to determine the presence of vascular structures in the graphic of means \(\mu _i\). Generally, between two consecutive peaks multiples valleys can appear. In this work, we select a unique valley v between two consecutive peaks to represent a unique candidate for a vessel. Then, we obtain the set \(V_{cand}\) with all valleys \(v_i\). Finally, we process the set \(V_{cand}\) by selecting the candidates that represent the vascular structures. For this, we will use as selection criterion the force of fall f of the valley \(v_i\) with respect to its two nearest peaks (see Fig. 5). The results of this approach can be observed in Fig. 6(b), (d).

Fig. 5.
figure 5

Adaptive threshold approach, where \(v_{i}\) represents an valley and \(p_{i}\) represents an peak. The valleys with force \(f_i \ge f\) are marked as vessels.

Full size image

3 Experimental Results

The proposed method was tested with 128 OCT histological sections, where 560 vascular structures have been manually labelled by an expert clinician. The images were taken with a confocal scanning laser ophthalmoscope, \(\text {CIRRUS}^\text {TM}\) HD-OCT Zeiss, with Spectral Domain Technology, at a resolution of \(490 \times 500\) pixels. Regarding the adaptive threshold approach, a force of fall, f, has been empirically established to a value of 50. We evaluated the accuracy of the proposed method using two metrics: precision and recall. Table 1 summarizes the results obtained by each approach. As we can see, the adaptive threshold approach offers a better and more complete performance.

Fig. 6.
figure 6

Results of the approaches. Red arrows, the vascular segments annotated by the experts. Red circles, the results of the different approaches. (a), (c) Fixed threshold approach. (b), (d) Adaptive threshold approach. (Color figure online)

Full size image
Table 1. Precision and recall results in the vessel detection process.
Full size table

4 Discussion and Conclusions

In this paper, new strategies for the automatic detection of retinal blood vessels in OCT images are presented. The proposed methodology exploits two approaches, a fixed and an adaptive threshold, over the mean intensity of all the columns between the retinal layers (RPE/C) and (M/E), to detect vascular structures. We used OCT images labelled by an expert clinician to evaluate the robustness and accuracy of the proposals, obtaining promising results. Analysing the results of the experiments (see Table 1), we can conclude that the adaptive threshold approach offers a more robust and coherent behaviour than the method with the fixed threshold. This is because the adaptive method is more robust to deviations in the global tendency of the column means, as explained before. An example of this situation is illustrated in Fig. 6(c), (d), where we can observe that the mean tendency moves darker at the right columns of the image, making that the fixed threshold approach produce several false positives in this region. However, the adaptive threshold approach, as there is no significant dropping profiles with respect to their surroundings, is capable to overcome this situation. Both proposals offer an automatic vessel identification, facilitating the work of the ophthalmologists in diagnostic processes of many vascular pathologies that are present in the retina facilitating, therefore, therapeutic processes.