Performance of the Brasiliano 92 orange cultivar with six trifoliate rootstocks

1.   Introduction

Globally, hundreds of millions of people suffer from eye diseases. Conditions such as diabetic retinopathy, cataracts, age-related macular degeneration, uncorrected refractive errors and glaucoma contribute significantly to visual impairment ^[1]. Among various fundus pathologies, changes in the morphology and structure of retinal vessels are frequently observed, carrying vital information about the overall health of the eye and the body. Hence, segmentation of the retinal vessels holds significant importance for the early screening, diagnosis and treatment of related diseases. Initially, specialists relied on the manual segmentation of retinal vessels, which is a laborious, monotonous and time-consuming task ^[2], particularly when dealing with a large number of segments. Therefore, the development of computer-aided diagnostic systems for the rapid and efficient automated segmentation of fundus vessels has become immensely important for ocular medical research and applications.

However, due to its own characteristics, there are a number of challenges in the retinal vessel segmentation task. First, retinal images often exhibit uneven brightness ^[3], resulting in low contrast between the blood vessels and the background, particularly in capillaries, which are difficult to discern visually. Second, the presence of hemorrhages, hard exudates and other interferences in the fundus retinal image, which are affected by the lesion, can cause significant noise in vessel segmentation ^[4]. Lastly, the retinal vasculature is highly structured. Unlike polyp and skin datasets, retinal vessels start from the fovea and gradually spread throughout the eye varying in thickness and overlapping with each other, which makes segmentation more difficult.

In the past, numerous researchers have dedicated their efforts to extracting the structure and morphology of the retina from fundus images. Currently, the primary segmentation methods can be categorized into two groups: traditional methods using hand-designed features and methods employing deep learning. Traditional segmentation methods, which extract the vascular structure directly by building algorithmic models, can be broadly classified into matched filtering methods ^[5], morphological ^[6] and mathematical modeling methods ^[7], vascular tracking methods ^[8] and variable model methods ^[9]. Although traditional methods have shown progress in retinal vessel segmentation, they heavily depend on manual feature extraction and parameter selection, which can be labor-intensive. Moreover, these methods yield poor segmentation accuracy for processes of capturing vessel endings and cannot meet the requirements of clinical practice.

With the continuous development of deep learning, the proposal of convolutional neural networks has led to impressive achievements in medical image segmentation. Fully convolutional networks (FCNs) ^[10] constitute the basis of the first method to use convolutional neural networks for semantic segmentation. Fu et al. ^[11] introduced a network model called DeepVessel which utilizes an FCN as a framework and incorporates multi-scale features and conditional random fields. At that time DeepVessel achieved remarkable results on tasks of retinal vessel segmentation. Following this, Ronneberger et al. ^[12] proposed a classical network known as U-Net, which demonstrated excellent feature extraction capability and outstanding performance. Since then, U-Net and its variants ^{[13,14,15,16]} are still the most popular segmentation models in medical image processing and are widely employed for medical image segmentation tasks, such as retinal vessel segmentation. Nevertheless, these methods exhibit certain limitations in attempts to achieve high-precision retina segmentation. These include a restricted receptive domain, the loss of detailed information caused by changes in feature map size and a decline in segmentation accuracy, as attributed to the semantic gap.

To solve the above problems, we propose an efficient and flexible multi-fusion network with grouped attention (MAG-Net) for accurate retinal vessel segmentation. First, we use the classical data preprocessing scheme proposed by Jiang et al. ^[17] to initially alleviate the problems of uneven brightness in fundus images. Second, we adopt a U-shaped encoder-decoder structure ^[18], which performs well on denoising tasks, and introduce depthwise separable convolution and atrous convolution into the encoder. These operations enable enlargement of receptive field without inflation of the parameters and ensure that the feature map can learn more long-range contextual information. Additionally, we introduce a grouped attention enhancement (GAE) module for skip connections to further enhance the traditional channel attention method. By using high-level features to complement low-level features, GAE aims to reduce the semantic gap between the encoder and decoder. Finally, to further improve the model's learning capability, we propose a multi-scale feature fusion (MFF) module. This module integrates features from different decoder levels to obtain a more comprehensive and precise feature representation, helping the image restoration.

Our primary contributions can be divided into the following points:

1) We propose a new multi-fusion network with grouped attention that enhances feature extraction capability and improves model accuracy and robustness.

2) We introduce a grouped attention fusion module that facilitates information interaction between the encoder and decoder, preserving valuable information across different levels of features.

3) We have evaluated the proposed model on publicly available datasets (DRIVE, CHASE and STARE), and the results demonstrate the model's strong segmentation performance and overall effectiveness.

2.   Related works

In recent years, U-Net and its variant models have demonstrated excellent segmentation capabilities on tasks of retinal vessel segmentation. The incorporation of innovative optimization mechanisms, including convolutional strategies ^[19], feature fusion techniques ^[20] and attention mechanisms ^[21], has further enhanced the segmentation performance of these models.

In traditional convolutional neural networks, the application of the receptive field plays a crucial role in segmentation ^[22]. How to improve the receptive field and enhance the feature extraction ability has been a popular research topic. DUNet ^[23] introduces deformable convolution, which adaptively adjusts the receptive field based on the scale and shape of the vessels. This adjustment allows the model to capture retinal vessels with diverse shapes and scales. SCS-Net ^[24] is a novel scale and context sensitive network for retinal vessel segmentation, and it includes a scale-aware feature aggregation module to dynamically adjust the receptive field to obtain multi-scale information. Zhang et al. ^[25] proposed a semi-isotropic model that maintains a relatively large receptive field at different stages of the convolutional network, demonstrating excellent segmentation performance. To enhance the feature extraction capability, Liu et al. proposed a ResDO-UNet ^[26] and applied it to retinal blood vessel segmentation by combining it with the DO-conv ^[27] layer. MCDAU-Net ^[28] was designed with a cascading dilated spatial pyramid pool module that enhances the receptive field and generates feature maps that are rich in contextual information.

In recent years, attention mechanisms have transcended their initial application of natural language processing and gained significant traction in medical image segmentation, especially on tasks of retinal vessel segmentation. For example, Guo et al. ^[21] proposed a lightweight network called SA-UNet, which incorporates channel attention to generate a spatial attention map of features along the spatial dimension for adaptive feature refinement. FANet ^[29] introduces a novel approach by integrating the current training features with the mask generated in the previous iteration to generate hard attention, which aims to suppress unwanted feature clutter. WA-Net ^[30] has been proposed as a width attention-based convolutional neural network that weighs the channel relevance via a width attention module, which considers the location and position correlations of feature maps. To overcome the difficulty of segmenting small blood vessels, a generator with attention augmented convolution has been proposed in ^[31] to highlight the region of interest in the whole image. DOLG-NeXt ^[32] integrates SE-Net and ConvNeXt, which overcomes the problem that the large transformer-based model performs poorly in small biomedical image environments, as well as realizes effective feature extraction and fine target segmentation.

While these methods have excelled on tasks of retinal vessel segmentation, they are not without limitations. Challenges such as the loss of fine-grained details during upsampling, significant computational overhead and the inclusion of excessive irrelevant information in attention mechanisms still persist.

3.   Methods

3.1.   Proposed network model

The complex structure of blood vessels in fundus retinal images, which are susceptible to uneven illumination and noise, poses challenges for retinal vessel segmentation. To improve the effectiveness of retinal vessel segmentation, we propose the MAG-Net, a multi-fusion network with grouped attention.

Figure 1 illustrates the overall network structure of the MAG-Net, which consists of three components: the encoder-decoder structure, GAE module and MFF module. The backbone of our model is based on the three-layer U-Net structure, renowned for its exceptional performance and versatility on diverse medical image segmentation tasks. Initially, the input image is preprocessed and divided into greyscale images of size 48 $\times$ 48, which are then fed into the encoder. To downsample the feature maps, a maximum pooling operation is applied between the encoder blocks, reducing the dimensions by half. Similarly, a transposed convolution operation is employed between the decoder blocks to double the size of the feature maps. The bottleneck is composed of two convolution operations followed by batch normalization and ReLU activation. The letter $C$ illustrates the changes in the feature maps for channels at each stage in Figure 1. In the MAG-Net, the hybrid convolutional fusion (HCF) module was first designed to obtain a larger receptive field and better feature extraction capability. Second, the GAE module is utilized to establish connections between corresponding encoder and decoder layers, facilitating the utilization of multi-scale features and enhancing the integration of detailed information from different locations. Finally, the MFF module combines feature maps generated at different scales in the decoder to better capture information at different levels and scales in the image. The three important modules are described in detail below.

Figure 1.  MAG-Net network structure. The letter $C$ indicates the number of channels of the current feature.

DownLoad: Full-Size Img PowerPoint

3.2.   Hybrid convolutional fusion module

To enhance the extraction capability of the encoder, the HCF module has been incorporated, and each original encoder module has been replaced by an HCF module. The HCF module combines depthwise separable convolutions and a spatial attention module, as illustrated in Figure 2. By employing depthwise convolutions with varying dilation rates, the HCF module encodes the fused features through element-level summation. Feature maps are then fed into the spatial attention module, making the network prioritize the vasculature. Next, a pointwise convolution is applied to perform linear transformations and nonlinear activations on individual pixels in each channel. Finally, we use a standard convolution operation, followed by DropBlock to prevent overfitting and accelerate the network convergence.

Figure 2.  Structure of the HCF module.

DownLoad: Full-Size Img PowerPoint

Spatial attention is a machine learning technique that is commonly used in image segmentation tasks; it emphasizes the interaction between features at different locations in spatial dimensions to help extract local features. The spatial attention module first performs global average pooling and global maximum pooling operations on the input features $F\in R^{H\times W\times C}$ along the channel dimension to obtain global information in the form of $F_{avg}^s\in R^{H\times W\times 1}$ and $F_{max}^s \in R^{H\times W\times 1}$, and it subsequently generates spatial attention weights $M^s\in R^{H\times W\times 1}$ applying a 7 × 7 convolution operation and a sigmoid activation function. Finally, the spatial attention weights are multiplied element-wise with the input features and aggregated to obtain the output features $F_s$, thus allowing the network to focus on features at different locations for enhancement. In summary, the output features $F_s$ can be expressed as

where $\delta$ denotes a sigmoid function, $f^{7\times 7}$ denotes a 7 × 7 convolution operation and $[; ]$ represents channel concatenation.

3.3.   Grouped attention enhancement module

The integration of skip connections in U-Net facilitates information exchange between the encoder and the decoder, preventing information loss and enhancing training efficiency. However, the presence of skip connections can lead to a semantic gap problem, as caused by information distortion and loss within the encoder and the decoder. To mitigate this issue, the GAE module has been introduced into the skip connection. This module incorporates feature maps from different levels and utilizes high-level features, which contain rich category information, to guide the low-level features. The goal is to achieve more precise segmentation of intricate details.

The channel attention module, which is often used to selectively generate importance weights for each channel, is an integral component of the GAE module. The module details are depicted in Figure 3. Assuming that $X_i$ denotes the input feature and $H$, $W$ and $C$ denote its height, width and number of input channels respectively. The channels are first compressed by using global average pooling $P_{avg}$ and global maximum pooling $P_{max}$ to obtain the outputs $P_{avg}(X_{i}) \in R^{1 \times 1 \times C}$ and $P_{max}(X_{i}) \in R^{1 \times 1 \times C}$. This approach allows the pooling pixels to be taken into account and the important features to be extracted. A multi-layer perceptron (MLP) is used to stimulate the channels and adaptively recalibrate the channel relationships. The MLP consists of two fully connected layers with the first fully connected layer having an output channel number of $C/r$, where $r = 4$. Subsequently, a ReLU activation function is introduced, and the output channel number of the second fully connected layer is restored to $C$. Subsequently, the stimulated results are fed into a sigmoid function and summed to obtain the channel attention weights $\beta_i$. The exact computational procedure can be expressed as follows:

Figure 3.  Structure of the channel attention module.

DownLoad: Full-Size Img PowerPoint

The specific structure of the GAE module is shown in Figure 4.

Figure 4.  Structure of the GAE module.

DownLoad: Full-Size Img PowerPoint

First, we use standard and transposed convolution operations to resize the low-level and high-level features in the encoder, respectively, to the same size and concatenate them into multi-scale features along the channel dimension. The multi-scale features are given as follows:

where $F_{low}$ indicates the low-level features and $F_{high}$ indicates the high-level features. Additionally, MF denotes the multi-scale features, $f_{3 \times 3}$ denotes a 3 $\times$ 3 standard convolution and $up()$ means a transposed convolution operation.

Second, we employ a channel attention module to extract the channel attention weights of the feature maps at different scales, denoted as $\beta_1$ and $\beta_2$, respectively. The detailed calculation will be explained in the next paragraph. The multi-scale channel attention weights are obtained by cascading, and they can be expressed as follows:

where $\beta$ is the multi-scale channel attention weight. Next, a softmax function is applied to $\beta$ to obtain the recalibrated soft attention $att_i$. $att_i$ contains the complete spatial information and the attention weights on the channel dimensions that facilitate the interaction between a particular feature channel attention and another feature channel attention. This operation is defined as follows:

$att$ is obtained by concatenating $att_i$ to denote the multi-scale channel attention weights after channel attention interactions, and it is expressed as follows:

Subsequently, the recalibrated multi-scale channel attention weight $att$ is multiplied by the multi-scale features to obtain the feature $F_c$. Finally, $F_c$ is the output after applying a 1 $\times$ 1 convolution operation, which rescales channels and sums with adjusted low-level features.

where $f_{1 \times 1}$ denotes a 1 $\times$ 1 convolution operation.

3.4.   Multi-scale feature fusion module

Due to the multivariate shape and structure of fundus vessels, existing vessel segmentation methods still exhibit limitations. To overcome the problem of feature adaptation for retinal vessels at different scales, we introduced the MFF module. This module leverages bilinear interpolation upsampling to complete the semantic interaction between neighboring low-level features and high-level features as well as the aggregation of features at different scales. It merges spatial and channel information, obtaining detailed local and global information. The structure of the MFF module is shown in Figure 5.

Figure 5.  Structure of the MFF module.

DownLoad: Full-Size Img PowerPoint

Specifically, the MFF module initiates at the bottleneck which is positioned between the encoder and decoder. It performs a level-by-level decoding process while employing a 1 $\times$ 1 convolution operation at different scales, reducing the number of feature channels to 16 to decrease the computational effort. Next, the MFF module conducts an operation by applying bottom-up bilinear interpolation along the decoder, thus merging with the adjacent upper-level feature maps via element-by-element summation. Thereafter, a nonlinear activation function, ReLU, is applied to improve the fitting of the nonlinear model and the nonlinear transformation. After upsampling and feature merging, the feature map achieves a flexible and adaptive feature transformation and incorporates differences in information between scales, complementing the spatial and channel information. In addition, the MFF module employs upsampling and cascades the activated features from each layer through bilinear interpolation to obtain multi-scale cascaded features (MCFs). Finally, through a series of convolution operations, batch normalization and nonlinear activation, the MCFs recombine the spatial and channel information to generate multi-scale aggregated features, which are cascaded with an output feature map along the decoder to obtain the final output.

4.   Datasets and evaluation metrics

4.1.   Datasets

We validated our method on three standard public datasets (DRIVE, CHASE, STARE). These datasets are described in detail as follows:

1) DRIVE ^[33]. This dataset contains 40 digital retinal images, namely, the corresponding groundtruth images and the corresponding masked images, 20 of which are used for training and the other 20 for testing. The size of each image is 565 $\times$ 584 pixels, and each image contains annotations of retinal regions and vascular regions.

2) CHASE ^[34]. This dataset contains 28 digital retinal images, namely, the corresponding groundtruth images and the corresponding masked images, 14 of which are normal retinal images, and the other 14 are retinal images with lesions. The size of each image is 999 $\times$ 960 pixels. Twenty images were set as the training set, and the other eight images were set as the test set.

3) STARE ^[35]. This dataset contains 20 fundus images, namely, the corresponding groundtruth images and the corresponding masked images with an image size of 605 $\times$ 700 pixels. We divided the dataset with the first 10 images used for training and the last 10 images used for testing.

4.2.   Evaluation indicators

The segmentation performance of retinal vessels was quantitatively evaluated in this study by using several metrics: Dice coefficient, accuracy, sensitivity and specificity. These metrics were assessed by using a confusion matrix. The Dice coefficient measures the similarity between the predicted and labeled results, while the accuracy represents the ratio of correctly segmented pixels to the total pixels providing an overall assessment of segmentation accuracy. Sensitivity evaluates the model's ability to segment the vascular region by measuring the proportion of correctly segmented positive samples. Specificity quantifies the model's capability to segment non-vascular regions by measuring the proportion of correctly segmented negative samples. Accuracy reflects the accuracy of the model's prediction as a positive sample. Each metric is defined as follows:

where true positive (TP) represents correctly segmented vessel pixels, true negative (TN) represents correctly segmented background pixels, false positive (FP) denotes the background pixels incorrectly segmented as vessel pixels and false negative (FN) denotes the vessel pixels incorrectly segmented as background pixels.

5.   Experiments and results

5.1.   Implementation details

Our deep learning method, implemented through the use of a PyTorch framework, was evaluated on an Ubuntu 64-bit operating system using a QuadroRTX 6000 server. During training, we employed a random patch approach with a patch size set to 48 $\times$ 48 pixels and a total number of 104800 patches. The model underwent 100 iterations with an initial learning rate of 0.001. For the DRIVE dataset and CHASE dataset, the training batch size was 128 and the threshold was 0.48. For the STARE dataset, we set the batch size to 64 and the threshold to 0.48. The model was optimized by using the Adam optimizer, with the exponential decay rate $\beta_1$ = 0.9, $\beta_2$ = 0.999 and the momentum $\epsilon$ = 1 $\times$ $10^{-8}$.

The loss function we use is the cross-entropy loss function, defined as follows:

where $y_i$ denotes the true label and $\widehat{y_i}$ represents the predicted label.

5.2.   Preprocessing

The problem of uneven brightness in the original retinal image significantly interferes with automatic segmentation. Preprocessing of the fundus images can effectively alleviate this problem. We have processed the original images by employing the following four steps and the effect of each preprocessing step was as shown in Figure 6.

Figure 6.  Preprocessing results: (a) original fundus vascular medicine image; (b) greyscale map; (c) data normalized image; (d) adaptive histogram equalized image; (e) gamma adjusted image.

DownLoad: Full-Size Img PowerPoint

1) Three-channel fusion of the color images to convert the corresponding greyscale images;

2) Normalization of greyscale images to mean and standard deviation to improve the convergence speed of the model;

3) Adaptive histogram equalization (CLAHE) of greyscale images to increase the contrast between the blood vessels and the background and suppress noise;

4) Gamma correction for vessel images to suppress illumination unevenness and centerline reflections.

5.3.   Module ablation

To verify the validity of each component in the MAG-Net, we conducted ablation experiments on the DRIVE, CHASE and STARE datasets. Specifically, we evaluated the impact of HCF, GAE and MFF, individually. Our baseline model was U-Net, while the standard model encompassed U-Net augmented with HCF, GAE and MFF. In the experimental setup, we also examined models without specific components, denoted as "w/o HCF", "w/o GAE" and "w/o MFF", respectively. The experimental results, displayed in Tables 1–3, highlight the most favorable outcomes for better visibility and analysis.

Table 1.  Ablation experiments on DRIVE dataset. AUC_ROC: area under the receiver operating characteristic curve, Acc: accuracy, Se: sensitivity and Dice: Dice coefficient.

Model	AUC_ROC	Sp	Se	Acc	Dice
baseline	0.9891	0.9856	0.8375	0.9704	0.8529
w/o HCF	0.989	0.9846	0.8474	0.9705	0.8549
w/o GAE	0.9892	0.9837	0.8544	0.9705	0.8556
w/o MFF	0.9891	0.9842	0.8473	0.9701	0.8532
ours	0.9895	0.9836	0.8588	0.9708	0.8576

---

 | Show Table

DownLoad: CSV

Table 2.  Ablation experiments on the CHASE dataset.

Model	AUC_ROC	Sp	Se	Acc	Dice
baseline	0.9867	0.9883	0.7876	0.9765	0.7970
w/o HCF	0.9878	0.9885	0.7957	0.9772	0.8031
w/o GAE	0.9878	0.9872	0.8104	0.9769	0.8040
w/o MFF	0.9880	0.9876	0.8014	0.9767	0.8009
ours	0.9884	0.9875	0.8123	0.9773	0.8069

---

 | Show Table

DownLoad: CSV

Table 3.  Ablation experiments on the STARE dataset.

Model	AUC_ROC	Pr	Sp	Se	Acc	Dice
baseline	0.9886	0.9166	0.9932	0.7461	0.9742	0.8161
w/o HCF	0.9895	0.9195	0.9929	0.7578	0.9748	0.8219
w/o GAE	0.9890	0.9138	0.9915	0.7604	0.9738	0.8167
w/o MFF	0.9882	0.9141	0.9915	0.7676	0.9743	0.8208
ours	0.9895	0.9158	0.9906	0.7781	0.9743	0.8228

---

 | Show Table

DownLoad: CSV

As can be seen in Tables 1–3, the MAG-Net achieved better overall performance compared than the baseline model. The area under the receiver operating characteristic curve (AUC_ROC), accuracy, sensitivity and Dice coefficient for the MAG-Net respectively reached 0.9895, 0.9708, 0.8588 and 0.8576 on the DRIVE dataset, 0.9884, 0.9773, 0.8123 and 0.8069 on the CHASE dataset and 0.9895, 0.7781, 0.9743 and 0.8228 on the STARE dataset. When we removed the HCF module from the standard model and replaced it with the standard convolution operation in U-Net, the model showed a slight decrease of 0.04 and 1.14% for the AUC_ROC and 0.06 and 1.66% for sensitivity on the DRIVE dataset and CHASE dataset, respectively, and a decrease of 2.01% for sensitivity on the STARE dataset. It demonstrates that increasing the perceptual field by using depthwise atrous convolutions with different dilation rates is effective in segmenting vascular pixels accurately. When we removed the GAE, the data in Tables 1–3 revealed that the absence of the GAE module has a small effect, as the effects were more pronounced for the accuracy and Dice coefficient, which decreased by 0.03 and 0.29%, 0.04 and 0.29% and 0.05 and 0.61%, respectively. Results emphasize the importance of the GAE module in the MAG-Net. The effectiveness of the MFF module is also well demonstrated by the fact that the three metrics, i.e., sensitivity, accuracy and the Dice coefficient, decreased by 1.15, 0.07 and 0.42% on the DRIVE dataset and 1.09, 0.06 and 0.60% on the CHASE dataset without the MFF.

In Figure 7, we show the segmentation results of various models in the ablation experiments on three datasets, aiming to visually validate the effectiveness of our module in segmenting retinal vessel details. The figure showcases six columns, (a) to (f), representing the preprocessed fundus images, ground truth annotations and the segmentation results obtained by using "w/o HCF", "w/o GAE", "w/o MFF" and the MAG-Net, respectively. The results on the DRIVE, CHASE and STARE datasets are shown from top to bottom in the figure above. Upon examination of the images, it is evident that all four models possess the ability to accurately segment the main trunk of the vessel and finer vessels across different datasets. Using capillary segmentation as a significant metric, the MAG-Net demonstrates its exceptional performance. We have highlighted the segmentation of some capillaries by marking and enlarging the same areas of each image with red boxes.

Figure 7.  Visualization of the results of the ablation experiments on the three datasets. From top to bottom: DRIVE dataset, CHASE dataset and STARE dataset are indicated respectively. From left to right: (a) preprocessed image, (b) groundtruth image, (c) w/o HCF, (d) w/o GAE, (e) w/o HFF and (f) MAG-Net.

DownLoad: Full-Size Img PowerPoint

5.4.   Experiments on the ablation of attention modules

The GAE module is a very important part of the MAG-Net, and to validate the reliability of our proposed grouped attention, we conducted a series of ablation experiments for the GAE module. We selected the widely used SE module ^[36], CBAM ^[37] module and GC block ^[38] to substitute the attention computation method in the GAE module and conducted experiments on the DRIVE dataset. Table 4 shows the results of our experiments.

Table 4.  Ablation experiments with different attention methods on the DRIVE dataset.

Model	AUC_ROC	Sp	Se	Acc	Dice
MAG-Net+SE	0.9892	0.9837	0.8517	0.9702	0.8542
MAG-Net+CBAM	0.9891	0.9850	0.844	0.9706	0.8547
MAG-Net+GC	0.9894	0.9849	0.8471	0.9708	0.8559
ours	0.9895	0.9836	0.8588	0.9708	0.8576

---

 | Show Table

DownLoad: CSV

Although the addition of the SE, CBAM and GC block improves the model's performance to a certain extent, these three methods are slightly inferior compared to the grouped attention approach. This is because the GAE module not only learns the channel weights of features at different scales, it also retains the details of low-level features and high-level features to the greatest extent possible. At the same time, the GAE module is simple in structure and achieves efficient segmentation with only a small amount of computation.

5.5.   Qualitative experiment

Figure 8 presents a visual comparison of our model, MAG-Net, with other competing models. We selected U-Net ^[12], SA-UNet ^[21] and Attention-UNet ^[39] as the competition, and we applied a test image from both the DRIVE and CHASE datasets for visual analysis, arranged in left-to-right order. Columns (a)–(f) display the preprocessed original image, the groundtruth image, U-Net segmented results, Attention-UNet segmented results, SA-UNet segmented results and MAG-Net segmented results, respectively. Under each image, we included an enlarged local area to promote the comparison of the results.

Figure 8.  Visualization results on the DRIVE dataset and CHASE dataset. From top to bottom: DRIVE dataset and CHASE dataset are indicated respectively. From left to right: (a) preprocessed image, (b) groundtruth image, (c) U-Net, (d) Attention U-Net, (e) SA-UNet and (f) MAG-Net.

DownLoad: Full-Size Img PowerPoint

By comparing the detailed regions, it can be seen that U-Net, Attention U-Net, SA-UNet and MAG-Net already have the ability to extract the vascular backbone from the original image and achieves satisfactory segmentation for most vascular regions. However, when comparing the groudtruth images, disconnections and mis-segmentations in the capillary region are easy to observe. Comparing the details, it can be seen that the MAG-Net does well at denoising, minimizes mis-segmentation and has excellent vessel connectivity.

5.6.   Quantitative experiment

To further demonstrate the superiority of our method, we compared the MAG-Net with some other deep learning-based retinal vessel segmentation methods. We evaluated them by using four metrics: specificity, sensitivity, accuracy and the Dice coefficient, respectively. Tables 5–7 indicate their segmentation results on the DRIVE, CHASE, and STARE datasets, respectively. When compared to existing segmentation methods, our proposed method showed significant improvements for most metrics, especially accuracy, outperforming the other methods on all three datasets. As can be seen in Table 5, the MAG-Net exhibits competitive ability that is not exhibited by other methods, with the best accuracy and Dice coefficient, which improved by 0.04 and 2.73%, respectively. As shown in Table 6, on the CHASE dataset, the MAG-Net excelled in the cases of the specificity and accuracy metrics, achieving 98.75 and 97.73% respectively. As can be seen in Table 7, similar to the CHASE dataset, the MAG-Net performed best for specificity and accuracy metrics, with a respective improvement of 0.28 and 0.18%.

Table 5.  Comparison with other methods on the DRIVE dataset.

Model	Year	Sp	Se	Acc	Dice
R2U-Net [40]	2018	0.9813	0.7792	0.9556	0.8171
DUNet [23]	2019	0.9800	0.7963	0.9566	0.8237
NFN+ [41]	2020	0.9813	0.7996	0.9582	0.8295
SA-UNet [21]	2020	0.9840	0.8212	0.9698	0.8263
MPS-Net [42]	2021	0.9740	0.8361	0.9563	0.8278
SCS-Net [24]	2021	0.9838	0.8289	0.9697	0.8189
AACA-MLA-D-UNet [43]	2022	0.9805	0.8046	0.9581	0.8303
Bridge-Net [44]	2022	0.9818	0.7853	0.9565	0.8203
CRAUNet [45]	2022	-	0.7954	0.9586	0.8302
SDDC-Net [46]	2023	0.9808	0.8603	0.9704	0.8289
ours	2023	0.9836	0.8588	0.9708	0.8576

---

 | Show Table

DownLoad: CSV

Table 6.  Comparison with other methods on the CHASE dataset.

Model	Year	Sp	Se	Acc	Dice
R2U-Net [40]	2018	0.9862	0.8298	0.9712	0.8475
DUNet [23]	2019	0.9752	0.8155	0.9610	0.7883
NFN+ [41]	2020	0.9880	0.8003	0.9688	0.8369
SA-UNet [21]	2020	0.9835	0.8573	0.9755	0.8153
SCS-Net [24]	2021	0.9839	0.8365	0.9744	-
MPS-Net [42]	2021	0.9795	0.8488	0.9668	0.8332
AACA-MLA-D-UNet [43]	2022	0.9801	0.8402	0.9673	0.8246
Bridge-Net [44]	2022	0.9840	0.8132	0.9667	0.8293
CRAUNet [45]	2022	-	0.8259	0.9659	0.8156
SDDC-Net [46]	2023	0.9789	0.8268	0.9669	0.7965
ours	2023	0.9875	0.8123	0.9773	0.8069

---

 | Show Table

DownLoad: CSV

Table 7.  Comparison with other methods on the STARE dataset.

Model	Year	Sp	Se	Acc	Dice
R2U-Net [40]	2018	0.9820	0.7756	0.9634	0.7928
DUNet [23]	2019	0.9878	0.7595	0.9641	0.8143
NFN+ [41]	2020	0.9863	0.7963	0.9672	0.8298
SCS-Net [24]	2021	0.9839	0.8207	0.9736	-
MPS-Net [42]	2021	0.9819	0.8566	0.9689	0.8491
AACA-MLA-D-UNet [43]	2022	0.9870	0.7914	0.9665	0.8276
Bridge-Net [44]	2022	0.9864	0.8002	0.9668	0.8289
SDDC-Net [46]	2023	0.9784	0.7988	0.9642	0.7776
ours	2023	0.9906	0.7781	0.9743	0.8228

---

 | Show Table

DownLoad: CSV

5.7.   Generalization ability of proposed multi-fusion network with grouped attention

Generalization ability is a very important performance evaluation metric for retinal vessel segmentation tasks. We evaluated the generalization ability of the model by performing cross-validation on different training and test sets. Table 8 gives a comparison of the generalization ability of the proposed method with three other methods, under the condition of the training being performed on the DRIVE dataset and testing performed on the STARE dataset, as well as the opposite experiment. Experiments show that our method has good overall performance and excellent performance on the Dice coefficient and sensitivity metrics, which shows that our method has excellent generalization ability in the task of vessel segmentation.

Table 8.  Cross-validation results.

Database	Methods	Sp	Se	Acc	Dice
STARE (trained on DRIVE)	MS-CANet [47]	-	-	0.9673	0.7826
	GDF-Net [48]	0.9795	0.7089	0.9588	-
	AA-WGAN [31]	0.9883	0.7839	0.9647	-
	ours	0.9745	0.8471	0.9648	0.7855
DRIVE (trained on STARE)	MS-CANet [47]	-	-	0.9701	0.8035
	GDF-Net [48]	0.9902	0.7289	0.9593	-
	AA-WGAN [31]	0.9928	0.8250	0.9635	-
	ours	0.9828	0.8255	0.9690	0.8236

---

 | Show Table

DownLoad: CSV

5.8.   Receiver operating characteristic curve evaluation and number of model parameters

In addition, we present the Receiver operating characteristic (ROC) curves for the different models in the ablation experiment for the three datasets. As can be seen in Figure 9, the values of the area under the curve for all four models were close to 1, indicating the models' superior performance for retinal vessel segmentation. Of these, the MAG-Net had the largest area of the ROC curve among all four models for all three datasets, further emphasizing the superiority of the MAG-Net over the other models in the three ablation experiments.

Figure 9.  ROC curves for the ablation models on different datasets.

DownLoad: Full-Size Img PowerPoint

Furthermore, we compared our work with other state of the art models in terms of the number of model parameters on the DRIVE dataset. The comparison results are shown in Table 9. It can be seen that our model demonstrated lower complexity than the other methods. In other words, our model performs well in terms of efficiency.

Table 9.  Model complexity comparison with state-of-the-art methods.

	U-Net	ResUnet	R2U-Net	AACA-MLA-D-UNet	MRC-Net [49]	ours
# Param.	4.32M	32.61M	39.09M	2.03M	0.9M	0.6M

---

 | Show Table

DownLoad: CSV

6.   Conclusions and discussion

This paper presents a new retinal vessel segmentation framework, the MAG-Net, which combines a multi-scale technique and an attention mechanism to realize improvements. The HCF module was introduced, and it utilizes spatial attention and multiple convolutions to expand the receptive field of the encoder, minimize noise and optimize the performance of the feature extractor at the source. The GAE module uses high-level features to guide low-level features on skip connections, optimizing segmentation details and alleviating the semantic gap between the encoder and decoder. The MFF module addresses the information loss problem during upsampling in the decoder by integrating features of multiple scales, which serves to supplement the detailed information. We validated our method on three datasets, i.e., the DRIVE, the CHASE and the STARE datasets. The experimental results show that our method achieves superior retinal vessel segmentation performance, as compared to U-Net, Attention-UNet and SA-UNet.

However, there are certain limitations of our work. First, the current implementation relies on annotated datasets, which can be expensive and restricts the availability of a larger number of annotated datasets. Second, our network was specifically designed for retinal vessel segmentation, while fundus images contain a wealth of information related to various fundus diseases. A single-target segmentation architecture is temporarily unable to meet the diagnostic needs of multiple fundus diseases.

Regarding future research, our research goals can be divided into two main points. First, we will focus on the improvement and optimization of the model to make it more adaptable to the specificity of fundus images and to increase its robustness for applicability to different pathological situations. Second, transfer learning and self-supervised learning are effective means to address data scarcity and labeling difficulties. In the absence of large-scale labeled data, we will consider exploring how to learn knowledge from other medical image tasks through transfer learning and use self-supervised learning to improve the performance of the model.

Use of AI tools declaration

The authors declare they have not used Artificial Intelligence (AI) tools in the creation of this article.

Acknowledgments

Our work was supported in part by the National Natural Science Foundation of China (No. 61962054), the National Natural Science Foundation of China (No. 61163036), the Cultivation Plan of Major Scientific Research Projects of Northwest Normal University (No. NWNU-LKZD2021-06).

Conflict of interest

We declare that there is no conflict of interest.