A fast and fully automated system for segmenting retinal blood vessels in fundus images

2.1 Image pre-processing stage

The primary objective of this stage is to enhance the retinal image quality by addressing non-uniform illumination, contrast normalization, and background noise reduction. The proposed image pre-processing pipeline consists of the following steps:

1. Green channel extraction: Isolating the green component from the color retinal image.

2. Illumination correction and contrast enhancement: Applying a logarithmic transformation to rectify uneven illumination and accentuate edge structures within the retinal image.

3. Noise reduction: Employing an improved dual-tree complex wavelet transform-based denoising technique to decompose the image and suppress noise through a shrinkage process.

4. Edge enhancement: Implementing anisotropic diffusion filtering to amplify the visibility of blood vessel edges in the RFI.

Apparently, human observers can identify the blood vessel structures in images of full-color contrast, though the proposed approach entails a computational maze through the three-color bandwidths, thereby protracting time in generating the final images with the segregated blood vessels. To counter this, the presented morphological filtering algorithm begins with the process of selecting the green channel of the input image. This channel offers superior differentiation between background and retinal blood vessels compared to other color channels, as illustrated in Figure 4c.

Figure 4

Comparative analysis of a color retinal image with its constituent color channels: (a) original image, (b) red channel, (c) green channel, and (d) blue channel. Images obtained from DRIVE dataset.

The steps followed by the segmentation algorithm are presented for an example image in Figure 5. Logarithmic transformation is a mathematical technique for enhancing contrast in retinal images [15]. This method effectively maps narrow pixel values with low intensities to a broad spectrum of output intensities. As a result, the dark areas of the image are stretched to greater intensity, resulting in a noticeable increase in the overall image contrast and brightness. When logarithmic functions are applied to pixel values, the method compresses high-intensity values, broadens low values, and thus highlights details that are normally not visible in contrast retinal images in a few cases are highlighted (Figure 5c). This technique has shown promise to improve visualization of vital structures in the retina and subsequently facilitates imaging analysis and used for diagnosis. A logarithmic transformation can be applied to the input image within the proposed system as follows:

Figure 5

Sequential overview of the proposed system’s output stages: (a) original input image, (b) isolated green channel, (c) image transformed by the logarithmic method, (d) result of applying the I-CWT, (e) output following adaptive AD filtering, (f) image after undergoing morphological operations, (g) final edge-detected image, and (h) superimposed visualization of the extent of the segmented image on the automated segmented image on the original retinal image. Images obtained from DRIVE dataset.

The modified output image, denoted as r″, is generated through the application of a logarithmic transformation to the green channel of the input image, represented by f. A constant value of 0.5 is assigned to parameter c. Logarithmic transformations are commonly employed to amplify the intensity levels within an image. Its particular efficacy lies in enhancing the visibility of details present in low-intensity regions. This technique is especially valuable for accentuating salient features when the image undergoes subsequent transformational processing.

Then, an I-CWT was carried out as a dependable image denoising method to denoise the retinal image without losing the details of blood vessels, including edges and curves (Figure 5d). Classical methods of denoising based on wavelets are known to have some disadvantages, for example, shift sensitivity, oscillatory artefacts, aliasing, and insufficient directionality [16]. In this work, an I-CWT, including a shrinkage process, was used to reduce the noise and enhance the weak patterns of the retinal blood vessel in the image. Wavelet shrinkage denoising is primarily based on thresholding wavelet coefficients, as for the core idea of wavelet-based shrinkage denoising. A lot of coefficients have small values, and several coefficients may contain very negligible values, which indicate the noises and small features of the images, whereas the coefficients with high amplitudes contain significant features of the images. In order to get enhanced visibility of the blood vessels, an AD filter [17] was applied as follows:

where t is the threshold value used to enable/disable a pixel, which can be optimized in the binary classification process (see Section 3.3), and λ is subject to 0 ≤ λ ≤ 1/4. In our study, the values for λ and t parameters are considered to be 0.20 and 5, respectively. It is equally important to look at the value of λ = 0. These values of N = 20 and t = 5 have been chosen after some empirical optimization done during the study. In this way, different models using these particular values were validated, among others, by testing them on other data splits, where the results showed that such values provided fairly good and steady performance in the case of binary classification tasks balancing precision and recall. Since λ belongs to the range of values 0 ≤ λ ≤ 1/4, such selection provides enough freedom in the model while still avoiding excessive regularization, which will help to retain the high efficiency of the algorithm. Furthermore, if we compare it with other possible values, then λ = 0. 20, t = 5 offered better performance in terms of high Acc and F1-scores without considering computational time. The rationale for choosing the examined parameter values in this research anchors on the empirical findings and theoretical analysis presented here. Additionally, I refers to the vessel blood sample, the c factors represent the transmission coefficients that are reorganized each repetition as a brightness gradient function, t is the restatement index, and ∇ denotes the near neighbor modifications in all the directions, north (N), south (S), east (E) and west (W), as follows:

The effect of this filtering is shown in Figure 5e, where the blood vessels appear smoother compared to that in Figure 5d, allowing for improved subsequent traceability.

2.2 Image segmentation stage

This phase is composed of two groups of operations. First, morphological operations to extract accurately the image components, and then the detection procedure of edges, which follows the following four steps: noise reduction, vessel segmentation, morphological refinement, and vessel linking.

3.1 DRIVE and HRF datasets

The efficacy of the proposed segmentation approach was evaluated using the publicly accessible DRIVE and HRF RFI datasets. Renowned for their associated GT images, meticulously annotated by multiple human experts, these datasets facilitate rigorous performance assessment. By comparing segmentation results against the GT, the system’s reliability and efficiency can be objectively measured. The objective of these datasets is to foster comparative analyses and advancements in automated blood vessel segmentation within retinal fundus imagery.

The DRIVE dataset [14] constitutes a collection of 40 color retinal images partitioned into training and testing subsets, each containing 20 images. Accompanying each image is a mask delineating the field of view (FOV) and a corresponding GT image. Expert annotations were provided for blood vessels in the training set images, which were subsequently utilized to optimize the proposed segmentation algorithms’ parameters. Conversely, the testing set’s vessel annotations were independently generated by two experts to ensure unbiased performance evaluation. Derived from a cohort of 400 diabetic patients aged 25–90 in the Netherlands, the dataset comprises 33 images without DR and 7 with mild early-stage indications. Taken with a Canon CR5 non-mydriatic 3-CCD camera with an FOV of 45°, the images are 8-bit color and are 768 × 584 pixels in TIFF (Tag Image File Format) format. Some original sample RFIs, along with the GT provided manually from the DRIVE dataset, are shown in Figure 6.

Figure 6

Sample of an image from the DRIVE dataset: (a) original fundus image, (b) FOV mask, (c) manual segmentation by expert 1, and (d) manual segmentation by expert 2.

The HRF dataset [19] comprises 45 RFIs obtained from three distinct patient cohorts: the lowest level of vascular endothelial growth factor expression was observed in healthy controls, intermediate in DR patients, and highest in glaucoma patients. Each group provided 15 monochromatic images obtained with a mydriatic CANON CF-60UVi fundus camera with a 60-degree FOV. These are 24-bit color JPEG images in the size of 3,504 × 2,336 pixels and the size of each pixel is 6.05 μm × 6.46 μm. In this study, the HRF dataset was used for evaluating the proposed model. Some sample images from the HRF dataset are depicted in Figure 7, along with their GT segmentations. For the comparison of the image within the region of interest, the FOV masks are circular binary masks (Figure 7b). All the images were manually segmented for blood vessels by three retinal image interpretation experts for each image.

Figure 7

Samples of the HRF dataset: (a) original image, (b) field of view (FOV) mask, and (c) ground truth segmentation done by human annotator.

3.2 Evaluation matrices

Using a binary classification, the input image is divided into vessel and non-vessel pixels by the proposed algorithm. If a pixel is correctly segmented as a vessel and also exists in the GT image, then it is said to be a true positive. On the other hand, the pixel that is detected by the algorithm as a vessel and is not marked as a vessel in the GT map is considered a false positive (Table 1). To assess the efficacy of the proposed algorithms, five standard quantitative metrics were computed: Acc, Sen, Spe, PPV, and NPV. These metrics are computed as averages across the dataset. The following nomenclature is employed in this work: TP for true positives, TN for true negatives, FP for false positives, and FN for false negatives.

Acc is defined as the percentage of correctly classified pixels to the total number of pixels within the FOV mask, given by

Sen is the ratio of TP to the total number of vessels given by

Spe is the ratio of TN to the total number of non-vessels pixels, expressed by

PPV is the proportion of TP pixels out of the whole referred pixels, given by

Finally, the NPV is the ratio of the TN pixels, which is the non-vessel pixels over the total number of pixels predicted as non-vessel, for instance, background. It can be calculated as

The interpretation of the confusion matrix, together with the employed performance metrics, is given in Table 1.

3.3 DRIVE dataset results

Initially, the efficacy of our segmentation system in identifying retinal blood vessels is assessed using the DRIVE dataset. Rigorous experiments were conducted on the training set images to optimize the parameter values for maximizing segmentation accuracy. For example, in the suggested morphological filtering approach (equation (2)), λ and t were established at 0.20 and 5, respectively. A 7 × 7 pixel Gaussian filter was employed, while the low and high thresholds were fixed at 25 and 40, respectively. A binary classification process was employed to categorize each pixel within the image as either a vessel or non-vessel based on a threshold value t. To optimize the threshold, a segmentation accuracy metric was calculated for a series of 100 experiments. In each iteration, the threshold value t was incremented by 0.1. The results for the evaluation metrics considered are given in Table 2 for the testing set of the DRIVE dataset.

Table 2 shows the qualitative performance of the proposed segmentation system to the testing set of the DRIVE dataset observed by two observers. The system achieves high results, demonstrating 98.67% Acc according to Observer 1 and 99.33% for Observer 2. This means that the system is very efficient in defining true retinal vasculature pixels of an image (av. 99%). The value for Sen, which in a way calculates the TP, is also good with a mean of 95.83%, but as we can see, it is somewhat lower than the Acc of the algorithm. This implies that although the system has high detection of blood vessels, there exists a possibility of some TP being missed.

Another positive result can be observed when it comes to the level of Spe, where there is an average of 98.62%. These results indicate that the system is effective at recognizing areas that are NOT part of the vessel. Observer 1 recorded 99.11% Spe, which is slightly higher than Observer 2 with 98.13%. In the case of PPV, the percentage of correctly identified vessel pixels, we can observe more variation between the observers with the average PPV of 91.34%. This value is not bad and implies that there is a moderate rate of FP for the given data. The NPV, which indicates the overall performance of the algorithm in correctly classifying non-vessel pixels, is also very high, with an average of 94%, thus offering great assurance in classifying non-vessel areas.

Overall, this research demonstrates that the system possesses favorable results in all the metrics for its performance, especially Acc and Spe since these are significant factors when it comes to segmentation in medical image applications. The trade-offs are slightly lower Sen and PPV, which points to the opportunities for optimizing its functioning and decreasing the number of FP and FN results. The differences between the observers mean that while the fully automated system was quite stable in the assessment of the video, it could be improved in its equality for different reliable evaluators. With the objective of evaluating our proposed vessel segmentation against several conventional methods, a comparison of results was performed and is shown in Table 3, using the DRIVE dataset.

The suggested system performs far better than all the competing methods in terms of almost all the assessment parameters, including Acc (99.33%) and Sen (96.27%). It also proves that we have very high accuracy of the proposed system for the classification of both vessel and non-vessel pixels. This is clearly evident from the outstanding Sen score, as depicted from the capability of the vessel pixels detection; hence, it has a low probability of omission on true vessel pixels.

Compared with Jin et al. [20] and other advanced methods, while the Spec (98.13%) of the proposed system is slightly lower than the rest of the methods compared, the proposed system is superior in other aspects. Here, the Spec score has also remained high so that the proposed system achieves a high level of possibility to detect the non-vessel pixels. The PPV of the proposed system, which is 92.56%, shows that it has a high degree of accuracy in detecting true vessel pixels from the vessel group. This is particularly important in diagnostics, where false alarms are massively detrimental since patients may be subjected to invasive procedures that are not required. Also, the NPV of 95.78% transmits that 95.78% of the negative cases or those that are not associated with the disease also accurately had negative tests. This value also enhances the confidence of the system in detecting the non-vessel pixels and hence reduces the tendency of the system to miss out on some pixels as wrong.

In comparison with the methods worked out at the present state of the art, like Keerthiveena et al. [27], where they have the highest value of 92.7% for Sen, our proposed system provides even higher results both in terms of Sen, Acc, and PPV. This means that the proposed system is fairer and stronger to address the problem of vessel segmentation than previous studies since it reduces comparative results on FP and FN cases. Due to the high accuracy and sensitivity of the proposed system, the suggested approach would be very beneficial where an extremely accurate and sensitive segmentation of the retinal vessels is needed, for example, in the clinical context. In diseases such as DR or macular degeneration, it is very important to be able to distinguish vessels with a relatively low error rate.

3.4 HRF dataset results

This section gives the performance assessment of the proposed vessel segmentation method utilizing the HRF database. The same parameter configuration optimized on the DRIVE dataset was employed for consistency. To assess the performance, the usual five standard evaluation metrics were computed by comparing the algorithm outputs to GT images provided within the HRF dataset. The resulting metrics are tabulated in Table 4.

Table 4 shows how the proposed segmentation method has a very high level of accuracy and can be considered robust. Specifically, with an Acc of 99.11%, the system works reliably in the detection of the retinal vessels, thus suggesting that it correctly classified most of the pixels contained in the images. A Sen of 97.97% underlines the ability of the system to identify TPs, meaning almost all real-life structures of vessels are detected. This is important in areas such as medicine, where the absence of vessel structures can result in wrong identification. Furthermore, a value of 98.87% for Spe comprises the precise recognition of the non-vessel regions, thus reducing wrong localization or excessive segmentation of the image. The PPV was 97.98%, giving the measure on how accurately detects the vessel pixels by the system, thus showing that in case the system recognizes a pixel as part of a vessel, then it will most probably be right. Last, an ideal NPV of 100% means that none of the pixels, which have been classified as non-vessel, actually contain vessels; thus, there is no possibility of FN results. Following the same strategy of comparison with DRIVE dataset, a comparative analysis of our blood vessel segmentation methods against contemporary state-of-the-art methods was conducted utilizing the HRF dataset. These results are tabulated in Table 5.

Based on Table 5, it is possible to conclude about high general performance level and the best results in Sen and NPV, which are both significant for medical image analysis. The suggested system yields an Acc of 99.11%, which is slightly lower than Kishore and Ananthamoorthy’s method (99.6%), showing that this method yields a high reliability of the classification of vessel and non-vessel pixels. In the case of Sen, the proposed system achieves the best result, which is 97.97%; that is, it can identify TPs at 97%, which means that almost all blood vessels are correctly segmented. This is very important in diagnosis because the next closest figure to the Sen score in the rest of the methods is much lower (72–79%). The system also has a very good Spe of 98.87%, having a correct identification of no-vessel areas with low FPs. Also, a PPV of 97 98% is higher than other compared methods and gives a good level of confidence in the identified vessel pixels into their respective classes. Its perfect NPV of 100% is redoing, especially in the aspect that it guarantees that none of the vessel pixels are obscured as background, which leads to highly fatal diagnostic negligence.

4.1 Vessel length

To measure the length of each retinal blood vessel segment (RBFVS), the skeleton of the vessel was first identified. Then, the total distance between the consecutive pixels within the segment was computed by the following equation:

where ( x i , y i ) define the coordinates of the sequence of pixels for the considered BVS, and N represents the total quantity of pixels.

4.2 Vessel density

The density of the retinal blood vessels was determined by calculating the percentage of the total surface area, which was occupied by the blood vessel pixels given by

where N _pix is the total sum of pixels for all the vessels presents in the image, and A _IM is the total area of the foveal image (mm²).

4.3 Vessel tortuosity

Its objective is to establish the average tortuosity coefficient of the total retinal vasculature. From the initial framework of the blood vessels, the branch points created in the vascular network were mapped to split the length of each BVS into a total number of b branch segments according to the following summation:

where s _i is the length of the ith branch segment within the BVS.

Therefore, the tortuosity BVS coefficient index (TC) is calculated as

where s length ( n ) is the nth vessel branch length (equation (13)) and s straight ( n ) , the straightforward distance between endpoints of the nth vessel branch segment, is estimated as follows:

where N is the number of fundamental pixels acquired from the branch of the blood vessels, and ( x , y ) are the pixel coordinates in each branch of the blood vessels. Finally, the average tortuosity coefficient of the whole blood vessel network was calculated using the mean of the tortuosity coefficients, which had been estimated for every blood vessel. ‎The equation represents the Euclidean distance between two points, specifically ( x 1 , y 1 ) and ( x N , y N ) , which are the coordinates of the first and last pixels in a blood vessel branch, respectively. The significance of this calculation lies in its contribution to determining the tortuosity coefficient, a measure of how much a blood vessel deviates from a straight path, calculated by comparing the actual curved length of the vessel ( s length ) , with its straight-line distance between the endpoints ( s straight ) . A perfectly straight vessel would yield a tortuosity coefficient of 1, as s length = s straight . However, as the vessel curves, s length becomes longer than s straight , resulting in a tortuosity coefficient greater than 1. The average tortuosity coefficient for the entire retinal vasculature is then calculated by averaging the tortuosity coefficients of all individual branches. This coefficient provides a valuable indicator of the overall curvature or tortuosity of the blood vessel network in the retina.

4.4 Vessel thickness

Retinal blood vessel thickness is thus defined as the mean diameter of the retinal vasculature in this work. A distance transform was performed on the inverse of the binary blood vessel image to determine the Euclidean distance of every vessel pixel to the nearest border. Pixels with maximum distance values were identified as the vessel centerlines. A tolerance was applied to account for floating-point computations. The average of all calculated distances represented the vessel's half-width, and the vessel thickness was determined by doubling this value.