Improved retinal vessel segmentation using the enhanced pre-processing method for high resolution fundus images

Introduction

Retinal images play a very important role in ensuring the early detection of symptoms relating to ocular diseases. Early detection will in turn enable timely treatment of eye diseases, which in most cases may significantly decrease the patients’ risk of total vision loss.¹ With the global prevalence of eye diseases being gradually on the rise annually, the World Health Organization has encouraged nations to have routine retinal screening in place.² This is intended to diagnose diseases such as diabetic retinopathy (DR), glaucoma and age-related macular degeneration (AMD) early enough so treatment can be administered before worse disease progression.³ Most hospitals are equipped with fundus cameras that can be used to generate fundus images by imaging a patient’s retina, samples of which are shown in Figure 1.

Figure 1. Samples of fundus images from a) Digital Retinal Images for Vessel Extraction (DRIVE) database and b) High-Resolution Fundus (HRF) database.

To assist ophthalmologists in performing efficient and accurate fundus image diagnosis, many studies have been conducted to automatically extract important parameters from a fundus image,^4–9 mainly focusing on automatic retinal blood vessel segmentation and then estimating vessel parameters from the segmentation output.^10–12 Figure 2 shows a typical flow of blood vessel segmentation procedure.

Figure 2. Processes involved in segmentation of retinal blood vessels from a fundus image (GCI = Green Channel Image, ROI = Region of Interest).

The blue-shaded box indicates the pre-processing steps focused in this study.

For validation of the blood vessel segmentation methods, most researches apply their methods to data from two popular benchmark databases, namely Digital Retinal Images for Vessel Extraction (DRIVE)¹³ and Structured Analysis of the Retina (STARE).¹⁴ However, it should be noted that these databases consist of images which are at a much lower resolution when compared to the fundus images produced by current modern fundus cameras. The images in the DRIVE database have resolution of 565 by 584 pixels while images in STARE have resolution of 700 by 605 pixels. This is because the databases date back to the early 2000s, and the fundus cameras of the time did not have the capability to produce high resolution images.

More recent studies have started to include databases with higher resolution fundus images, such as the High-Resolution Fundus (HRF) database.¹⁵ The images in the HRF database have resolution of 3504 by 2336 pixels, markedly higher than images in the DRIVE and STARE databases. Figure 1b shows a sample image from the HRF database.

From Figure 1, it can be seen that the region of interest (ROI) in the DRIVE image is surrounded by dark pixels. However, this is not the case with the image from the HRF database, Figure 1b. The top and bottom edges of the image are not surrounded by the dark area as in Figure 1a. This may result in noisy vessel segmentation output with false positive vessel pixels near the top and bottom images. To the best of our knowledge, this specific problem has not been addressed in the literature.

In this study, we investigated a simple and efficient way to eliminate the noisy pixels in segmentation output for fundus images whose ROIs are not fully surrounded by dark pixels, such as the case with images in HRF database. The proposed method is to be applied in the pre-processing step of retinal blood vessel segmentation workflow, illustrated as the shaded blue box in Figure 2. In order to validate the effectiveness of the proposed pre-processing step, vessel segmentation procedure based on an adapted Bar-Combination Of Shifted FIlter REsponses (B-COSFIRE) filter that we previously published¹⁶ is performed on the pre-processed output.

Methods

Pre-processing is one of the key steps in retinal blood vessel segmentation techniques, which helps to ensure that the initial fundus image is optimised for the subsequent vessel detection phase. The original red green blue (RGB) format of digital fundus images is not the optimal form for the accurate detection of retinal blood vessels from an image processing point of view due to the natural colours in fundus images that poorly contrast with the retinal background vessels. Issues such as inconsistent illumination across the image, lesser contrast between retinal blood vessels and the retinal background as well as noisy images are other concerns that need to be addressed during the pre-processing step, so that the input image for the vessel segmentation step will be of better clarity in terms of retinal blood vessel structures.

In this study, the pre-processing method employed by Soares¹⁷ is used as the basis since it is considered the established method for this purpose.¹⁸ Figure 3 illustrates the overview of the pre-processing steps where the first step is to extract the green channel image (GCI) from the color fundus image. The GCI displays a noticeably better vessel appearance, while the red channel image shows low vessel-to-background contrast and the blue channel image displays low dynamic range making the vessels appear almost invisible. This decision to use only GCI is supported by most previously established methods used for segmenting retinal blood vessels from fundus images.^18–21 The original code for pre-processing and vessel segmentation using B-COSFIRE can be obtained here.

Figure 3. Workflow of pre-processing steps applied to a fundus image prior to retinal blood vessel segmentation procedure (GCI = Green Channel Image, ROI = Region of Interest).

Results & discussion

As described in the introduction section, the modified B-COSFIRE¹⁶ filter is used to extract the vessel features from the pre-processed output images. Sample outputs of the pre-processed images with their corresponding vessel feature images using the single padding and double padding pre-processing methods are displayed in Table 1. By observing the HRF feature image produced using single padding in Table 1, dark lines can be seen on both the top and bottom borders of the vessel feature image. These dark lines will be highly likely to be segmented as false vessel pixels when processed for segmentation. However, the vessel feature image produced using the double padding method does not have these dark lines, thus decreasing the possibility of having a large number of false positive vessel pixels.

Table 1. Sample outputs of pre-processing steps applied to High-Resolution Fundus (HRF) images and the corresponding vessel feature images obtained using the Bar- Combination Of Shifted FIlter REsponses (B-COSFIRE) filter.

Another visible improvement is in terms of the brightness level of vessel pixels in the double-padded vessel feature image as opposed to single-padded. To confirm that double padding is better than single padding for the purpose of vessel segmentation, another comparison is performed on segmentation results using the different padding methods.

Table 2 shows segmentation outputs using the different padding methods, together with their zoomed-in versions. Apart from the apparent improvement in much reduced noisy pixels on top and lower border of the ROI, subtle improvements in vessel appearance are also observed. In general, double padding helps in further enhancing the vessel features, with most vessels appearing brighter compared to single padding outputs, including the smaller vessels. It is found that using double padding in pro-processing results in the successful removal of false positive pixels near the top and bottom image borders of all 45 segmentation output images in the HRF database. In order to quantify the improvement of the segmentation performance when using the proposed method, Table 3 summarises the performance metrics for segmentation using the different padding methods. Following the metric selection in our previous study, four metrics are included, namely Sensitivity (Sn), Specificity (Sp), Balanced Accuracy (B-Acc) and Matthew’s Correlation Coefficient (MCC).

Table 2. Comparison of quantitative segmentation results using single and double padding for pre-processing.

As expected, the application of double padding results in improved segmentation performance where performance values are increased across all the four considered metrics. This is attributed to the successful removal of the noisy pixels near the top and bottom borders of all images in HRF database, thus decreasing the number of false positive pixels and improving the overall segmentation performance.

Conclusion

In this study, we proposed a simple but effective method to improve blood vessel segmentation performance for images with the ROI reaching the image borders. The simple method of adding additional layer of dark pixels around the image proves to be effective in removing the noisy pixels at the image borders. Quantitatively, the additional padding step also managed to improve all the four considered metrics for vessel segmentation, namely Sensitivity (73.76%), Specificity (97.53%), Balanced-Accuracy (85.64%) and MCC value (71.57%) for the HRF database. This method has only been validated on a single high-resolution fundus image database for now, so in the future more databases should be included for validation to attest the robustness of the proposed methods on multiple databases. The proposed improvement method, while simplistic in nature, could prove to be very effective in increasing overall vessel segmentation performance, particularly for images that are not fully surrounded by dark pixels such as HRF database images.

Data availability