Vascular-aware mixture-of-experts with a texture-enhanced decoder for accurate X-ray vessel segmentation

Abstract

Background

Accurate vessel segmentation in X-ray angiography (XRA) is essential for diagnosing coronary artery disease and supporting interventional procedures. However, low contrast, complex backgrounds, and substantial variation in vessel caliber make accurate segmentation challenging, especially for thin vessels.

Methods

We propose a Vascular-Aware Mixture-of-Experts (VA-MoE) framework for XRA vessel segmentation. The model incorporates Dynamic Snake Convolution (DSnC) to capture the elongated and curved morphology of vessels and a Texture-Enhanced Decoder (TED) to recover fine spatial details and suppress background noise during upsampling. The proposed framework was evaluated on three representative datasets: ARCADE, DCA1, and XCAD.

Results

VA-MoE achieved an IoU of 87.80% and a Dice coefficient of 93.08% on ARCADE, an IoU of 65.18% and a Dice coefficient of 78.03% on DCA1, and an IoU of 67.52% and a Dice coefficient of 79.93% on XCAD. Compared with the strongest baseline methods, VA-MoE improved IoU and Dice by 2.77 and 1.21 percentage points on ARCADE, by 2.40 and 2.10 percentage points on DCA1, and by 2.89 and 1.50 percentage points on XCAD, respectively. Qualitative results further suggested improved topological continuity and better preservation of thin peripheral vessels compared with competing methods.

Conclusion

The proposed VA-MoE framework provides a promising solution for XRA vessel segmentation. By jointly enhancing vascular structural modeling and texture recovery, it may support automated cardiovascular image analysis.

Background

Segmenting coronary arteries from X-ray angiography (XRA) images is pivotal in aiding clinicians for the precise diagnosis and management of arterial diseases [1, 2]. It provides an essential basis for evaluating the degree of coronary artery stenosis, which is critical for determining the severity of the disease and guiding treatment decisions [3]. However, the inherent complexity of XRA imaging, combined with the intricacies of coronary vascular structures, presents significant challenges in accurately segmenting coronary arteries.

As shown in Fig. 1, the unique characteristics of XRA, such as the high variability in vessel appearance and prevalent image noise, complicate the task of delineating vascular topology with precision. These factors introduce substantial difficulties in achieving consistent and reliable segmentation, making it an ongoing challenge. To address this, traditional methods [4, 5] primarily rely on manually extracted features for vessel segmentation. While these approaches can accurately delineate high-contrast vascular regions, they often struggle with complex, interconnected areas. This limitation leads to fragmentation in the segmentation results, which can subsequently interfere with accurate clinical decision-making by healthcare professionals.

Fig. 1.

Challenges of vessel segmentation in X-ray angiography. Diverse morphologies and complex background noise complicate vascular delineation

To enhance the robustness and accuracy, numerous deep learning-based approaches have been applied to coronary artery segmentation. Existing methods can be further categorized into CNN-based, Transformer-based, and Mamba-based approaches. Among them, CNN-based methods are predominantly built upon the U-Net architecture [6] and enhance segmentation performance by incorporating multi-scale feature aggregation strategies [7, 8] and attention mechanism [9–11]. Compared with traditional algorithms, these methods demonstrate more stable and robust performance in coronary arteries segmentation. However, despite their effectiveness in local texture and multi-scale feature extraction, CNN-based methods may still require additional structural designs to model long-range vascular topology and complex vessel continuity.

Vision Transformer (ViT)-based methods, first introduced by Dosovitskiy et al. [12], provide an alternative mechanism for modeling long-range dependencies and global context. Inspired by this, Transformer-based variants [13–15] are introduced into coronary arteries segmentation. These methods leverage self-attention mechanism to effectively model the global context of coronary arteries and explicitly enhance the relational semantics among vessels, which is particularly beneficial for handling large morphological variations and complex vascular topologies. However, their substantial computational and memory overhead results in high hardware requirements. The emergence of Mamba-based methods [16–18] enables deep learning models to achieve an effective balance between CNNs and Transformers. Despite the notable progress achieved by the aforementioned methods, most of them primarily focus on strengthening the encoder to capture richer semantic representations, while the decoder is often treated as a simple upsampling and feature fusion module. For X-ray angiography (XRA) vessel segmentation, the decoder plays a critical role in recovering fine-grained vascular details, as it directly determines whether thin and low-contrast vessels can be successfully reconstructed from noisy and ambiguous feature maps. Thus, without explicit structural guidance, decoder features are easily dominated by texture noise and imaging artifacts, leading to discontinuous vessel predictions.

Recently, Mixture-of-Experts (MoE) architectures have demonstrated effectiveness in enhancing model capacity and specialization. Existing MoE designs [19–21] are not well suited for the vascular segmentation tasks. They typically employ generic convolutional experts and emphasize semantic diversity, without considering vessel-specific structural priors such as directional continuity, curvature adaptability, and topology preservation. Consequently, their potential for modeling complex vascular morphologies remains largely underexplored in XRA scenarios.

Motivated by these observations, we argue that accurate coronary artery segmentation requires a structure-aware decoding paradigm that explicitly incorporates vascular priors while remaining robust to severe imaging noise. To this end, we propose a novel framework that introduces vascular-aware expert modeling and texture-enhanced decoding in a unified manner. Specifically, we design a Vascular-Aware Mixture-of-Experts (VA-MoE) module that leverages dynamically adaptive snake convolutions to specialize experts in capturing vessel curvature and directional continuity. By generating stage-wise routing schemes, the proposed VA-MoE injects structural priors into the decoding process progressively. Furthermore, to suppress vessel-irrelevant noise and enhance fine-grained texture recovery, we introduce a Texture-Enhanced Decoder (TED) equipped with a parallel guidance strategy and a spatial-channel cooperative optimizer. This decoder effectively integrates structure-enhanced features from the VA-MoE module while preserving low-level texture cues, thereby improving the continuity and completeness of thin and low-contrast vessels.

In summary, the main contributions of this work are fourfold:

• We propose a novel vascular-aware mixture-of-experts framework that explicitly models coronary artery structures by incorporating vessel-specific priors into expert routing and feature decoding.

• We design a dynamic snake convolution–based expert architecture to effectively capture the elongated, curved, and topologically continuous nature of coronary vessels.

• We introduce a Texture-Enhanced Decoder that robustly suppresses noise and recovers fine vessel details through parallel structural guidance and spatial–channel cooperative optimization.

• Extensive experiments on multiple public X-ray coronary angiography datasets show that the proposed method consistently achieves favorable performance compared with state-of-the-art approaches in both segmentation accuracy and fine-vessel detection performance.

Related work

In this section, we review representative deep learning–based methods for coronary artery segmentation in X-ray angiography (XRA) images, as well as recent advances in Mixture-of-Experts (MoE) learning for medical image segmentation. Specifically, we first summarize existing XRA vessel segmentation approaches, which mainly focus on enhancing feature representation through convolutional, Transformer-based, and Mamba-based models. We then review MoE-based segmentation methods and discuss their limitations when applied to vascular structure modeling.

X-ray vessel segmentation methods

Early deep learning–based approaches for X-ray coronary artery segmentation are predominantly built upon convolutional neural networks (CNNs), which leverage local receptive fields to capture vascular appearance patterns. Representative methods typically enhance segmentation performance through multi-scale feature aggregation and attention mechanisms. For example, Zhang et al. [11] proposed CBAM-UNet, which incorporates channel and spatial attention modules into U-Net to strengthen the model’s focus on vascular regions and improve the extraction of fine coronary artery structures. Zhou et al. [7] introduced UNet++, employing nested skip connections and deep supervision to progressively integrate multi-scale features, thereby enabling more effective representation of thin vessels. In addition, Challier et al. [22] proposed CM-UNet, a contrastive masked U-Net framework that combines masked image reconstruction and contrastive feature learning for self-supervised pre-training, demonstrating improved segmentation performance under limited-annotation settings. Although these CNN-based methods achieve notable improvements in robustness and accuracy, their predominantly local operations may make explicit long-range topological modeling less straightforward, particularly for coronary arteries that are elongated, tortuous, and highly connected.

To address this limitation, Transformer- and Mamba–based methods have been introduced to enhance global context modeling in XRA vessel segmentation. Cao et al. [15] proposed Swin-Unet, which integrates Swin Transformer blocks into a U-shaped architecture. By leveraging shifted-window self-attention and multi-scale feature aggregation, Swin-Unet effectively captures long-range dependencies and improves the segmentation of complex vascular structures. More recently, Huang et al. [23] proposed SAM-VMNet, which integrates the Segment Anything Model with a vessel-aware Mamba U-Net. By combining SAM-generated coarse masks with domain-specific refinement, this method achieves accurate and robust coronary artery segmentation and stenosis quantification.

Despite the enhanced global modeling capability provided by self-attention mechanisms and state-space modeling, these approaches remain vulnerable to the intrinsic noise and low contrast of XRA images. In particular, the absence of explicit structural guidance during the decoding process often leads to unstable segmentation of thin or low-contrast vessels, resulting in fragmented predictions and missed branches.

Mixture-of-experts in medical image segmentation

Mixture-of-Experts (MoE) architectures have been explored in computer vision as an effective means to increase model capacity and enable expert specialization. For example, Wu et al. [24] proposed MoE-NuSeg, a two-stage mixture-of-experts framework for nuclei segmentation that employs multiple specialized experts to jointly model nuclei regions, boundaries, and clustered edges, thereby improving segmentation accuracy while simultaneously reducing model parameters and computational cost. In a different line of work, Li et al. [25] introduced MoE-SAM, which enhances SAM adaptation by incorporating a mixture-of-experts module to selectively refine and fuse multi-layer features, leading to stronger feature representations and improved medical image segmentation performance. Focusing on the non-forgetting adaptation of foundation models, Wang et al. [26] developed SAM-Med3D-MoE, a gated mixture-of-experts framework that integrates general and task-specific models without updating their original parameters, effectively mitigating catastrophic forgetting and significantly boosting 3D medical image segmentation accuracy. Although these methods achieve strong performance in medical image segmentation, they fail to consider the intrinsic structural characteristics of blood vessels.

Thus, we aim to extend the MoE paradigm towards a structure-aware formulation tailored for vascular segmentation. Instead of relying on generic expert designs and semantically driven routing, our approach emphasizes vessel-specific expert specialization by explicitly incorporating structural priors into the expert modeling process.

Methods

The proposed model consists of an encoder, a Texture-Enhanced Decoder (TED), and a Vascular-Aware Mixture-of-Experts (VA-MoE) module, as shown in Fig. 2.

Fig. 2.

Overview of the proposed VA-MoE framework. The pipeline consists of an encoder, the VA-MoE module, and the Texture-Enhanced Decoder

Encoder

The encoder is primarily responsible for encoding deep vascular features and providing structural and semantic guidance for both the VA-MoE module and the decoder. In this work, based on ResNet-34 [27], we introduced several minor modifications to tailor the architecture to our specific task. Considering the relatively small image size and the presence of substantial background noise, we replaced the original large-kernel convolution with a 3 × 3 convolution. This modification enables more effective extraction of vascular features while mitigating the influence of background noise.

We adopt the modified ResNet-34 as the feature extractor, which can achieve a good balance between contextual modeling and computational efficiency.

Specifically, the encoder consists of five encoding stages. Given an input image X ∈ R^H×W×3, it is first processed by a 3 × 3 convolution, producing an intermediate feature map X₁:

1

Subsequently, a max-pooling operation is applied to X1, reducing its spatial resolution by a factor of 2. A standard ResNet-34 basic block, consisting of two 3 × 3 convolutional layers with batch normalization, ReLU activation, and a residual shortcut, is then used to obtain the output of the second encoding stage, denoted as X2.

2

where RES(·) denotes the ResNet block and X′1 denotes the downsampled feature of X1. Subsequently, X2 is downsampled, and the same procedure is repeated to derive the remaining encoder outputs, including X3, X4, and X5.

Note that the first ResNet block does not employ downsampling. Finally, we preserve the feature representations from all encoder stages, forming a feature set {X_i | _i ∈ {1, 2, 3, 4, 5}}, which is utilized for subsequent vessel structure enhancement and decoding.

Vascular-aware mixture-of-experts module

The proposed Vascular-Aware Mixture-of-Experts (VA-MoE) module plays a critical role in modeling vascular topology and introducing structural priors into the decoder. Considering the highly tortuous, elongated, and spatially unpredictable nature of blood vessels, we adapt a task-adaptive mixture-of-experts routing strategy to vascular structure-aware expert selection. This design enables the decoder to receive stage-specific vascular guidance. Unlike the prior skip-connection routing design [28], the experts in the proposed VA-MoE module are constructed using DSnC to model vessel curvature and directional continuity, and depthwise separable convolutions are used after expert aggregation to control computational overhead.

Specifically, as shown in Fig. 3, we replace the original convolution operations in the conventional MoE with dynamic snake convolutions (DSnC) [29], enabling the VA-MoE to adaptively perceive vessel morphology and extract vascular structural features. Depthwise separable convolutions are further employed to reduce the additional computational cost introduced by expert modeling.

Fig. 3.

Structure of the VA-MoE module. The illustration corresponds to one decoding stage

Given the multi-level encoder features {Xi | i = 1, 2, 3, 4}, we first resize them to a unified spatial resolution by bilinear interpolation. The resized features are then concatenated and processed by a lightweight feature extractor to obtain an integrated feature representation X′. In this implementation, the lightweight feature extractor (·) is implemented as a 3 × 3 depthwise convolution:

3

where Resize(·) denotes bilinear interpolation, Agg(·) denotes feature aggregation by concatenation, and (·) denotes the 3 × 3 depthwise convolution used as the lightweight feature extractor. Based on X′, the VA-MoE module generates layer-specific routing schemes for each decoder stage. A set of gating functions {gi} is used to compute the expert-selection probabilities:

4

5

where Wi is a learnable weight matrix for the i-th expert and G represents the Top-K routing probabilities over N experts. In each decoding stage, the routing probabilities are generated from the structure-enhanced representation X′ and used to select the most relevant experts. To reduce redundancy, the same expert set is shared across stages, whereas the routing probabilities are stage-specific. This design allows the model to adaptively emphasize experts that better match local vessel curvature, directionality, and branching patterns. Next, to capture vessel curvature and directional continuity, each expert network is constructed using a combination of DSnC operations:

6

where DSnCx and DSnCy denote DSnC operations along two orthogonal spatial directions, ◦ represents sequential composition, and BN denotes batch normalization. In the DSnC-based expert, the offset field is generated from the input feature map through a lightweight convolutional branch. With a kernel size of 9, sampling points are progressively generated along two orthogonal snake-like directions from the kernel center, and the local displacement is bounded within [− 1, 1]. Bilinear interpolation is used for fractional sampling locations. This constrained x- and y-directional DSnC design enables the experts to model vessel elongation and curvature while reducing unstable deformation away from thin vessels. The outputs of the selected experts are then weighted by the routing probabilities and aggregated to obtain structure-enhanced features:

7

where Gi denotes the routing weight assigned to the i-th selected expert. The Top-K strategy activates only the most relevant experts for each stage, thereby controlling computational overhead and avoiding unnecessary expert redundancy. As the decoding process proceeds, the above operations are repeatedly applied at each decoder stage. Finally, the VA-MoE output is upsampled and refined by a lightweight convolution:

8

where DW(·) denotes depthwise separable convolution and Resize(·) represents bilinear upsampling. The resulting features are injected into the corresponding decoder layers to enhance vascular structure decoding.

Texture-enhanced decoder

X-ray angiography (XRA) images usually suffer from severe noise and imaging artifacts, which substantially hinder the decoder from accurately recovering fine-grained texture details. To address this issue, we propose a Texture-Enhanced Decoder that enhances the effective integration of VA-MoE features by adopting a parallel guidance strategy. In addition, a spatial-channel cooperative optimizer is embedded into each decoding stage to suppress vessel-irrelevant noise and artifacts in XRA images.

We denote the four outputs of the VA-MoE module as {V1, V2, V3, V4}, corresponding to different decoder stages. Given the encoder output X5, it is first upsampled to generate X′. Subsequently, based on the parallel guidance strategy, X ′, together with the resampled V₃ and V₄, are concatenated and processed by the spatial-channel cooperative optimizer to produce the structure-enhanced decoding feature D₄:

9

where Re(·) is resize operation, Concat(·) is the concatenation operation, and Optimizer(·) is spatial–channel cooperative optimizer, which is used to further suppress noise and enhance vessel-related textures. Specifically, as shown in Fig. 4, spatial attention (SA) [30] and channel attention (CA) [31] in the optimizer are applied in parallel to capture complementary contextual cues. The optimized feature is obtained by aggregating the two attention-enhanced responses:

10

Fig. 4.

Structure of the spatial-channel cooperative optimizer

where SA(·) and CA(·) denote the spatial attention and channel attention operations, respectively. By recursively applying the above fusion and optimization process across all decoding stages, the final decoder output is progressively refined.

By jointly integrating low-level texture cues and the structured representations provided by VA-MoE, the proposed Texture-Enhanced Decoder enables more accurate and robust decoding performance for the vascular segmentation task.

The rationale for this design is that XRA vessel reconstruction requires both structural continuity and local texture recovery. The parallel guidance strategy introduces VA-MoE-derived structural priors into multiple decoder stages instead of relying on a single late fusion operation, which helps maintain continuity for elongated and low-contrast vessels. Meanwhile, the spatial-channel cooperative optimizer suppresses vessel-irrelevant background responses in the spatial dimension and recalibrates vessel-sensitive channels in the channel dimension. These two operations are complementary for preserving thin branches while reducing noise-related false responses.

Loss function

To measure the discrepancy between the model predictions and the ground truth, we adopt a hybrid loss function that combines the Binary Cross-Entropy (BCE) loss and the Dice loss. The BCE loss is particularly effective at supervising pixel-wise classification and stabilizing the optimization process in vessel segmentation, especially under class-imbalance conditions. In contrast, the Dice loss explicitly optimizes region-level overlap, which helps preserve the continuity and completeness of thin and elongated structures. Therefore, the overall loss function of our framework is formulated as follows:

11

where pred and gt are the prediction and ground truth, respectively.

Experimental design

Datasets

To comprehensively evaluate the effectiveness and generalization capability of the proposed method, we conduct experiments on three publicly available X-ray coronary angiography (XRA) datasets, including ARCADE, DCA1, and XCAD, as shown in Fig. 5.

Fig. 5.

Representative examples from the DCA1, XCAD, and ARCADE datasets

ARCADE [32]. ARCADE is a large-scale dataset specifically collected for vascular lesion detection in X-ray coronary angiography. Since we focus on X-ray vascular segmentation, we select the SYNTAX subset of the ARCADE dataset as the segmentation dataset for subsequent evaluation. The images in the ARCADE dataset were collected using two different angiographic systems at the Research Institute of Cardiology and Internal Diseases, Almaty, Kazakhstan. The corresponding ground truth annotations were manually delineated by experienced cardiology experts using the Computer Vision Annotation Tool (CVAT). The provided .json annotation files were subsequently converted into binary mask representations for use in our experiments. Following the original split protocol of this dataset, we divide the data into training and test sets, resulting in 1000 images for training, and 200 images for testing.

DCA1 [33]. The DCA1 dataset is a public benchmark dataset specifically designed for X-ray vascular segmentation. It was collected by the Department of Cardiology of the Mexican Social Security Institute, with appropriate ethical approval. Specifically, the DCA1 dataset consists of 134 X-ray coronary angiography images, along with corresponding ground truth annotations delineated by experienced cardiology experts. In this paper, the entire dataset is divided into two subsets according to an 8:2 ratio, with 107 X-ray images used for training and 27 X-ray images used for testing.

XCAD [2]. The public XCAD dataset is the first X-ray angiography coronary artery disease dataset. It contains 1,621 mask frames, 1,621 coronary angiography images, and 126 coronary angiography images with corresponding ground-truth annotations labeled by experts. All images have a resolution of 512 × 512. Since our study focuses on fully supervised learning, we mainly use the 126 labeled images for training and testing. We split the dataset in the 8:2 ratio, resulting in a training set with 100 samples and a test set with 26 samples.

Experimental settings

All experiments were conducted on a Linux server running Ubuntu 18.04.6 LTS. The system was equipped with dual Intel Xeon Gold 6226R CPUs @ 2.90 GHz and 502 GB of system memory. GPU acceleration was provided by an NVIDIA GeForce RTX 3090, with NVIDIA driver version 535.98 and CUDA Runtime version 12.2. Regarding the software platform, all experiments were conducted based on U-Bench Zoo [34].

For hyperparameter settings, the batch size was set to 8, and the total number of training epochs was fixed at 300. The original XRA images were single-channel grayscale angiograms. For compatibility with the ResNet-based encoder, each image was replicated into three channels before being fed into the network. Data augmentation included random 90° rotations and random horizontal and vertical flips, whereas normalization was applied as a preprocessing step. To ensure a fair comparison, the same preprocessing and augmentation pipeline was consistently applied across all experiments. The models were optimized using stochastic gradient descent (SGD) with an initial learning rate of 0.01 and a weight decay of 1e-4. A polynomial learning-rate decay schedule was adopted during training. Following the unified loss configuration of U-Bench Zoo [34], the loss-weight hyperparameter λ was set to 0.5 for the BCE term. The same unified training protocol and loss setting were applied to all competing methods and the proposed model.

Evaluation metrics

To quantitatively evaluate the segmentation performance of the proposed method and competitors on three public X-ray vascular datasets, we adopt four widely used metrics in segmentation tasks, including Intersection over Union (IoU), Dice coefficient (Dice), Sensitivity (Sen), and Accuracy (Acc).

Specifically, IoU quantifies the region-level overlap between predictions and ground truth. Dice coefficient measures the overall similarity between the predicted segmentation and the ground truth. Sensitivity evaluates the detection rate of vascular pixels, which is critical for capturing thin, elongated structures. Accuracy assesses the overall global pixel-wise classification correctness. Because vessel segmentation is highly imbalanced at the pixel level, Acc was retained mainly for comparability with the existing benchmark metrics in the main tables and was not interpreted as a standalone indicator of model quality. Specificity, precision, and balanced accuracy were not added to the main comparison tables because complete and consistently verified pixel-level confusion-matrix results were not prepared for all competing methods under the same evaluation protocol. Their definitions are as follows:

12

where TP, TN, FP and FN are true positive, true negative, false positive and false negative pixels, respectively.

Results

Comparison with state-of-the-art methods

To systematically evaluate the effectiveness of the proposed method on X-ray coronary artery segmentation, we conducted comprehensive comparisons on three public datasets, including ARCADE (SYNTAX subset) [32], DCA1 [33], and XCAD [2]. We compared our method with multiple recent state-of-the-art segmentation approaches, including CFM-UNet [35], CSCA-UNet [36], CSWin-UNet [37], ERDUNet [38], GHUNet [39], MDSA-UNet [40], Perspective-UNet [41], and UNetV2 [42].

All competing methods were trained and evaluated under identical preprocessing, augmentation, and training settings to ensure a fair comparison. Therefore, the values in Tables 1, 2, 3 and 4 should be interpreted as within-study retraining results under a unified protocol, rather than as direct replacements for the originally reported values in the corresponding publications, which may have used different datasets, splits, preprocessing strategies, or implementation settings. Considering that coronary vessels are thin and elongated with complex branching structures, we emphasize Sensitivity (Sen) as a key metric for assessing the capability of detecting fine vessels, together with IoU, Dice and Acc.

Table 1.

Performance comparison on the ARCADE dataset

Methods	Year	IoU (%)↑	Dice (%)↑	Sen (%)↑	Acc (%)↑
Proposed	—	87.80	93.08	89.53	99.16
CFM-UNet [35]	2025	75.52	85.55	85.78	98.08
CSCA-UNet [36]	2024	82.37	89.58	86.91	98.43
CSWin-UNet [37]	2025	74.10	82.56	82.61	98.72
ERDUNet [38]	2024	82.52	90.06	87.09	98.63
GHUNet [39]	2025	79.58	91.87	88.54	98.65
MDSA-UNet [40]	2025	81.38	88.52	87.68	98.46
Perspective-UNet [41]	2023	85.03	91.46	89.22	98.61
UNetV2 [42]	2025	71.51	82.76	86.09	98.28

Legend: Higher values indicate better performance. IoU, intersection over union; Dice, Dice coefficient; Sen, sensitivity; Acc, accuracy

Table 2.

Performance comparison on the DCA1 dataset

Methods	Year	IoU (%)↑	Dice (%)↑	Sen (%)↑	Acc (%)↑
Proposed	—	65.18	78.03	83.08	97.63
CFM-UNet [35]	2025	55.30	69.88	75.09	96.77
CSCA-UNet [36]	2024	60.47	74.82	74.45	97.41
CSWin-UNet [37]	2025	50.23	66.06	67.91	95.89
ERDUNet [38]	2024	61.79	75.90	79.30	97.40
GHUNet [39]	2025	62.78	75.93	80.15	97.47
MDSA-UNet [40]	2025	59.19	73.57	76.18	97.20
Perspective-UNet [41]	2023	62.27	74.68	77.06	97.56
UNetV2 [42]	2025	49.59	65.93	66.39	96.43

Legend: Higher values indicate better performance. IoU, intersection over union; Dice, Dice coefficient; Sen, sensitivity; Acc, accuracy

Table 3.

Performance comparison on the XCAD dataset

Methods	Year	IoU (%)↑	Dice (%)↑	Sen (%)↑	Acc (%)↑
Proposed	—	67.52	79.93	80.84	97.83
CFM-UNet [35]	2025	56.48	71.88	74.69	96.82
CSCA-UNet [36]	2024	62.98	76.68	76.50	97.51
CSWin-UNet [37]	2025	50.40	63.83	67.30	96.35
ERDUNet [38]	2024	62.88	76.73	78.40	97.44
GHUNet [39]	2025	62.49	76.64	78.51	97.38
MDSA-UNet [40]	2025	59.49	74.00	76.97	97.07
Perspective-UNet [41]	2023	64.63	78.43	77.34	97.64
UNetV2 [42]	2025	47.86	62.77	64.06	96.19

Legend: Higher values indicate better performance. IoU, intersection over union; Dice, Dice coefficient; Sen, sensitivity; Acc, accuracy

Table 4.

Computational efficiency comparison of all models

Method	Year	#Param ↓	FPS ↑
CFM-UNet [35]	2025	57.20	34.86
CSCA-UNet [36]	2024	35.27	36.70
CSWin-UNet [37]	2025	23.57	17.91
ERDUNet [38]	2024	10.30	53.91
GHUNet [39]	2025	12.81	8.27
MDSA-UNet [40]	2025	6.58	61.62
Perspective-UNet [41]	2023	112.66	21.44
UNetV2 [42]	2025	25.15	33.65
Proposed	—	23.34	24.78

Legend: #Param denotes the number of parameters (M), where smaller values are preferred; FPS denotes inference speed, where larger values are preferred

Evaluation on the ARCADE dataset. As summarized in Table 1, the proposed method achieved the best performance across all evaluation metrics on the ARCADE dataset. Specifically, our approach attained an IoU of 87.80%, a Dice coefficient of 93.08%, a sensitivity of 89.53%, and an accuracy of 99.16%. Compared with the strongest baseline for each metric, the proposed method improved IoU by 2.77% points over Perspective-UNet, Dice by 1.21% points over GHUNet, sensitivity by 0.31% points over Perspective-UNet, and accuracy by 0.44% points over CSWin-UNet. These results suggest improved region overlap, boundary delineation, and fine-vessel detection on ARCADE.

While competing approaches tend to generate discontinuities or miss subtle vessel segments, the qualitative results suggest that the proposed method better preserves vessel continuity and recovers thinner low-contrast branches. Overall, these findings support the favorable performance of the proposed method on the ARCADE dataset.

Evaluation on the DCA1 dataset. As reported in Table 2, our method achieved the best overall performance across all evaluation metrics on the DCA1 dataset. It attained an IoU of 65.18%, a Dice coefficient of 78.03%, a sensitivity of 83.08%, and an accuracy of 97.63%. Compared with the strongest baseline for each metric, the proposed method improved IoU by 2.40% points over GHUNet, Dice by 2.10% points over GHUNet, sensitivity by 2.93% points over GHUNet, and accuracy by 0.07% points over Perspective-UNet. Because the DCA1 dataset contains a relatively small number of labeled images, these gains should be interpreted with appropriate caution and further validated in larger cohorts.

Evaluation on the XCAD dataset. Table 3 presents the quantitative comparisons on the XCAD dataset. Our method achieved the best performance across all evaluation metrics, obtaining an IoU of 67.52%, a Dice coefficient of 79.93%, a sensitivity of 80.84%, and an accuracy of 97.83%. Compared with the strongest baseline for each metric, the proposed method improved IoU by 2.89% points over Perspective-UNet, Dice by 1.50% points over Perspective-UNet, sensitivity by 2.33% points over GHUNet, and accuracy by 0.19% points over Perspective-UNet. These results indicate consistent improvements on XCAD, although the small size of this dataset requires cautious interpretation. Qualitative results are presented in Fig. 6.

Fig. 6.

Qualitative comparison of vessel segmentation results. Each row shows a representative sample, from left to right: image, ground truth, MDSA-UNet, Perspective-UNet, CSCA-UNet, and the proposed method

Efficiency analysis

Beyond segmentation accuracy, we further compare the computational efficiency of all models in terms of the number of parameters and inference speed, as shown in Table 4. Our method contains 23.34 M parameters and achieves an inference speed of 24.78 FPS, suggesting a reasonable trade-off between segmentation performance and computational cost. Although this speed may be acceptable for workstation- or GPU-supported auxiliary analysis, further compression, pruning, distillation, or lightweight optimization would be required for deployment in mobile, embedded, or resource-limited clinical environments.

Overall, our model maintains moderate computational efficiency while achieving favorable segmentation performance, particularly in sensitivity (Sen). These performance gains primarily stem from two key design choices. First, the proposed framework strengthens long-range structural modeling through vascular-aware expert routing. Second, the Texture-Enhanced Decoder improves the recovery of fine vessel details during upsampling. The balance between segmentation accuracy and deployment efficiency should be further evaluated in future clinical implementation studies.

Ablation studies

To quantitatively evaluate the contribution of each component, we conduct ablation experiments on the XCAD dataset under the same training and evaluation protocol as the main experiments. The evaluated components include the Texture-Enhanced Decoder (TED), the VA-MoE module, and the parallel guidance strategy (PGS). As summarized in Table 5, the progressive introduction of TED, VA-MoE, and PGS consistently improves segmentation performance over the baseline. The reported ablation results correspond to a single fixed training and evaluation setting for each configuration under the same data split and protocol. These ablation experiments were designed to assess the contribution of individual architectural components and hyperparameter choices rather than to perform formal statistical testing. Therefore, the results are reported descriptively, with emphasis on the direction and consistency of performance changes across metrics.

Table 5.

Component ablation on the XCAD dataset

Backbone	TED	VA-MoE	PGS	IoU(%)↑	Sen(%)↑	Acc(%)↑
✓				48.11	67.48	95.92
✓	✓			66.69	79.90	97.78
✓	✓	✓		67.19	80.42	97.81
✓	✓	✓	✓	67.52	80.84	97.83

Legend: TED denotes the Texture-Enhanced Decoder and PGS denotes the parallel guidance strategy. Check marks indicate included components

We first adopt a plain backbone followed by direct upsampling for mask prediction. This baseline achieves an IoU of 48.11%, a Sen of 67.48%, and an Acc of 95.92%, indicating limited capability in capturing fine vessel structures. By introducing the TED, the IoU improves from 48.11% to 66.69% (+ 18.58%), Sen increases from 67.48% to 79.90% (+ 12.42%), and Acc rises from 95.92% to 97.78% (+ 1.86%). This improvement indicates the value of enhanced multi-scale feature interaction and directional connectivity modeling. Further incorporating the VA-MoE module leads to additional improvements, with IoU increasing to 67.19% (+ 0.50%), Sen improving to 80.42% (+ 0.52%), and Acc reaching 97.81% (+ 0.03%). These results suggest that long-range dependency modeling provided by VA-MoE can further enhance the detection of thin and elongated vessels. Finally, enabling PGS yields the best performance, achieving an IoU of 67.52%, a Sen of 80.84%, and an Acc of 97.83%. Compared with the previous configuration, PGS contributes an additional + 0.33% IoU, + 0.42% Sen, and + 0.02% Acc, suggesting that parallel guidance during training may help stabilize optimization and enhance structural consistency. Overall, the ablation results support the contribution of each component, with TED providing the largest performance gain, VA-MoE further strengthening long-range structural modeling, and PGS refining structural continuity during training.

To further investigate the effectiveness of the proposed decoder, we visualize the feature responses with and without the decoder across three public datasets, as shown in Fig. 7. Without the proposed decoder (column (c)), the feature responses exhibit fragmented activations, weak continuity, and incomplete coverage of thin vessel branches. In particular, fine and low-contrast vessels are prone to missing or appear discontinuous, indicating insufficient global structural modeling. In contrast, after introducing the proposed decoder (column (d)), the feature activations become more concentrated along vessel regions, showing improved structural coherence and enhanced continuity.

Fig. 7.

Ablation visualization of the proposed decoder

Each row corresponds to one representative sample. From left to right: original image, ground truth, heatmap response without the proposed decoder, and heatmap response with the proposed decoder.

Thin branches are more clearly highlighted, and spurious background responses are suppressed. These observations suggest that the proposed decoder strengthens multi-scale feature aggregation and directional connectivity modeling, thereby improving vessel integrity and reducing missed branches in the qualitative examples.

To investigate the influence of the number of experts in the VA-MoE module, we conduct ablation experiments by varying the number of experts from 1 to 4 on the XCAD dataset, as shown in Table 6. When only a single expert is used, the model achieves an IoU of 66.84%, a Dice of 78.75%, a Sen of 79.28%, and an Acc of 97.69%. Increasing the number of experts to two leads to the best overall performance, with IoU improving to 67.52% (+ 0.68%), Dice to 79.93% (+ 1.18%), Sen to 80.84% (+ 1.56%), and Acc to 97.83% (+ 0.14%).

Table 6.

Expert-number ablation on the XCAD dataset

Number	IoU(%)↑	Dice(%)↑	Sen(%)↑	Acc(%)↑
1	66.84	78.75	79.28	97.69
2	67.52	79.93	80.84	97.83
3	67.31	79.42	80.35	97.73
4	66.62	78.47	79.07	97.64

Legend: Comparison of different numbers of experts in the VA-MoE module on the XCAD dataset

These results indicate that introducing a second expert may enhance feature diversity and improve long-range dependency modeling. However, further increasing the number of experts to three slightly degrades performance, with IoU decreasing to 67.31% (-0.21%), Dice to 79.42% (-0.51%), Sen to 80.35% (-0.49%), and Acc to 97.73% (-0.10%). A possible explanation is that excessive experts introduce redundant representations and increase optimization difficulty under limited training data, particularly on the small XCAD dataset. Therefore, the two-expert setting was selected as a balanced configuration for subsequent experiments.

To investigate the impact of the expert architecture within the VA-MoE module, we replace the proposed Dynamic Snake Convolution (DSnC) combination with three alternative designs: standard 3 × 3 convolution, dilated convolution (rate = 3), and depthwise separable convolution (DW Conv). The results are summarized in Table 7. When using standard 3 × 3 convolution, the model achieves an IoU of 66.79%, a Dice of 78.96%, a Sen of 79.88%, and an Acc of 97.71%, indicating stable local feature modeling capability. Replacing it with dilated convolution (rate = 3) slightly improves performance, yielding an IoU of 67.15% (+ 0.36%), a Dice of 79.58% (+ 0.62%), a Sen of 80.61% (+ 0.73%), and an Acc of 97.78% (+ 0.07%). This suggests that enlarging the receptive field benefits long-range vessel dependency modeling. In contrast, adopting depthwise separable convolution leads to noticeable performance degradation, with IoU dropping to 64.96%, Dice to 78.62%, Sen to 77.50%, and Acc to 97.69%. The reduced representational capacity of depthwise convolution may limit the modeling of complex vessel morphology. The proposed DSnC-based expert achieves the best overall performance in this ablation setting, reaching an IoU of 67.52%, a Dice of 79.93%, a Sen of 80.84%, and an Acc of 97.83%. Compared with standard 3 × 3 convolution, DSnC improves IoU, Dice, Sen, and Acc by + 0.73%, + 0.97%, + 0.96%, and + 0.12%, respectively. It also outperforms dilated convolution by + 0.37%, + 0.35%, + 0.23%, and + 0.05% across the four metrics. These results suggest that the DSnC combination is better suited to curved and elongated vessel structures in the present setting and provides a favorable balance between structural flexibility and representational capacity.

Table 7.

Expert-type ablation on the XCAD dataset

Method	IoU(%)↑	Dice(%)↑	Sen(%)↑	Acc(%)↑
3 × 3 Conv	66.79	78.96	79.88	97.71
Dilated Conv (rate = 3)	67.15	79.58	80.61	97.78
DW Conv	64.96	78.62	77.50	97.69
Proposed	67.52	79.93	80.84	97.83

Legend: DW Conv denotes depthwise separable convolution

To investigate the impact of injecting VA-MoE output features into different decoder stages, we conduct ablation experiments by concatenating the VA-MoE outputs into 1, 2, 3, and 4 decoding layers, respectively. The results are summarized in Table 8. When the VA-MoE features are injected into only one decoder layer, the model achieves an IoU of 66.94%, a Dice of 79.08%, a Sen of 79.97%, and an Acc of 97.75%, indicating that single-layer feature enhancement already contributes to improved structural modeling. Injecting the VA-MoE features into two decoder layers yields the best overall performance, with IoU increasing to 67.52% (+ 0.58%), Dice to 79.93% (+ 0.85%), Sen to 80.84% (+ 0.87%), and Acc to 97.83% (+ 0.08%). These results suggest that feature fusion at two decoding stages enables more effective structural information propagation while maintaining stable feature representation and optimization. However, when the number of injection layers increases to three, performance begins to decline, with IoU decreasing to 66.73% (-0.79%), Dice to 78.86% (-1.07%), Sen to 79.62% (-1.22%), and Acc to 97.60% (-0.23%) compared to the two-layer configuration. Further increasing the injection depth to four layers results in additional degradation, yielding an IoU of 66.47%, a Dice of 78.57%, a Sen of 79.43%, and an Acc of 97.54%. Therefore, injecting VA-MoE output features into two decoder stages achieves the best balance between feature representation capacity and information redundancy. Excessive multi-stage injection may introduce redundant information or interfere with decoder learning, thereby leading to marginal performance degradation.

Table 8.

Parallel-guidance ablation on the XCAD dataset

Parallel branch	IoU(%)↑	Dice(%)↑	Sen(%)↑	Acc(%)↑
1	66.94	79.08	79.97	97.75
2	67.52	79.93	80.84	97.83
3	66.73	78.86	79.62	97.60
4	66.47	78.57	79.43	97.54

Legend: Comparison of different numbers of parallel-guidance branches on the XCAD dataset

To further assess the consistency of the performance gains, we conducted an exploratory method-level statistical analysis based on the aggregated results reported in Tables 1, 2 and 3. The detailed results are summarized in Table 9. For each dataset and each metric, the absolute improvement of the proposed method over each of the eight baseline methods was calculated. The proposed method outperformed all baseline methods across IoU, Dice, sensitivity, and accuracy on ARCADE, DCA1, and XCAD. Across the baseline methods, the mean improvements were 8.80, 5.28, 2.79, and 0.68% points for IoU, Dice, sensitivity, and accuracy on ARCADE; 7.48, 5.93, 8.51, and 0.61% points on DCA1; and 9.12, 7.31, 6.62, and 0.78% points on XCAD, respectively. One-sided Wilcoxon signed-rank tests across baseline methods showed consistent positive improvements for all metrics on all three datasets (p = 0.0039). One-sided sign tests yielded the same p value because the proposed method outperformed all eight baseline methods for each metric. The SD and 95% CI in Table 9 summarize variability across baseline-method-level improvements and should not be interpreted as run-to-run variability or image-level uncertainty. Because this analysis was based on method-level aggregated results rather than image-level paired scores, it should be interpreted as exploratory and not as a substitute for image-level paired statistical testing.

Table 9.

Exploratory method-level improvement analysis across baseline methods

Dataset	Metric	n	Wins	Mean Δ	SD	95% CI	Wilcoxon p	Sign p
ARCADE	IoU	8	8/8	+ 8.80	4.76	4.82–12.77	0.0039	0.0039
ARCADE	Dice	8	8/8	+ 5.28	3.72	2.18–8.39	0.0039	0.0039
ARCADE	Sen	8	8/8	+ 2.79	2.03	1.09–4.49	0.0039	0.0039
ARCADE	Acc	8	8/8	+ 0.68	0.22	0.50–0.86	0.0039	0.0039
DCA1	IoU	8	8/8	+ 7.48	5.35	3.00-11.95	0.0039	0.0039
DCA1	Dice	8	8/8	+ 5.93	4.22	2.40–9.46	0.0039	0.0039
DCA1	Sen	8	8/8	+ 8.51	4.98	4.35–12.68	0.0039	0.0039
DCA1	Acc	8	8/8	+ 0.61	0.60	0.11–1.11	0.0039	0.0039
XCAD	IoU	8	8/8	+ 9.12	6.29	3.86–14.37	0.0039	0.0039
XCAD	Dice	8	8/8	+ 7.31	6.09	2.21–12.41	0.0039	0.0039
XCAD	Sen	8	8/8	+ 6.62	5.47	2.04–11.19	0.0039	0.0039
XCAD	Acc	8	8/8	+ 0.78	0.55	0.32–1.24	0.0039	0.0039

Legend: Δ denotes the absolute improvement of the proposed method over each baseline method. The analysis was performed across baseline methods using the aggregated values reported in Tables 1, 2 and 3. One-sided Wilcoxon signed-rank tests and one-sided sign tests were used to evaluate whether the improvements were consistently greater than zero. This exploratory method-level analysis was based on aggregated results rather than image-level paired scores and should not be interpreted as a substitute for image-level paired statistical testing. The SD and 95% CI summarize variability across baseline-method-level improvements and do not represent run-to-run variability or image-level uncertainty

Discussion

In this work, we proposed a vascular-aware segmentation framework for coronary artery segmentation in X-ray angiography (XRA) images to address the challenges posed by complex vessel morphology, severe imaging noise, and low contrast. Unlike approaches that mainly strengthen encoder representations, the proposed VA-MoE and TED emphasize vessel-specific structural priors during decoding. The results across ARCADE, DCA1, and XCAD suggest that combining vascular-aware expert routing with texture-enhanced decoding can improve region overlap, thin-vessel preservation, and structural continuity. Nevertheless, the observed advantages should be interpreted with caution because the comparisons are based on fixed dataset splits and aggregated performance metrics.

Recent foundation-model-based approaches have shown strong potential for medical image segmentation. For example, MedSAM extends the Segment Anything Model to medical images [43], adapter-based SAM variants improve domain adaptation for medical segmentation [44], and diffusion-based frameworks such as MedSegDiff-V2 provide an alternative generative paradigm for segmentation [45]. These approaches offer advantages in general representation learning, transferability, and prompt-based segmentation. However, XRA vessel segmentation remains challenging for foundation models because thin low-contrast vessels require precise boundary recovery, topological continuity, and robustness to background artifacts. In this context, the proposed VA-MoE framework may be complementary to foundation models: foundation models can provide general visual priors or coarse masks, whereas VA-MoE and TED can serve as vessel-specific refinement modules to enhance curvature adaptability and fine-vessel continuity.

The clinical relevance of segmentation improvements should also be interpreted carefully. Although Dice and IoU are widely used for evaluating segmentation overlap, higher overlap scores do not necessarily guarantee greater clinical utility. Alidoost et al. showed that the utility of a deep learning segmentation model may not be Dice coefficient dependent in a vessel segmentation setting [46]. For coronary XRA analysis, clinically meaningful evaluation may require radiologist-based scoring, vessel continuity assessment, stenosis visibility analysis, and downstream assessment of interventional decision support. Because the present study used publicly available anonymized datasets, additional clinician-based scoring, multi-reader evaluation, stenosed-region-specific analysis, and stenosis quantification remain beyond the scope of the present methodological study and should be incorporated in future clinically oriented validation. This is an important limitation and a direction for future work. In addition, specificity, precision, and balanced accuracy were not reported in the main comparison because complete pixel-level confusion matrices or prediction masks were not available for all competing methods; future studies should include these false-positive-sensitive metrics when complete outputs are available.

Several limitations should be acknowledged. First, the ground-truth masks in public datasets are based on manual expert annotations, which may contain boundary uncertainty, annotation errors, or inter-rater variability. Therefore, the model should be regarded as an auxiliary tool for reducing the burden of vessel delineation rather than as a method that exceeds expert judgment. Future studies should incorporate multi-expert annotations and inter-rater variability analysis to better characterize the upper bound and uncertainty of segmentation performance.

Second, the smaller DCA1 and XCAD datasets showed lower absolute segmentation performance than ARCADE. This may be related to limited labeled samples, differences in image quality and resolution, annotation uncertainty, data imbalance, or domain shift across imaging systems. Although data augmentation, depthwise separable convolution, and a fixed evaluation protocol were used to reduce overfitting risk, the present results cannot fully exclude overfitting under small-sample conditions. Future work should include larger multi-center datasets, repeated runs with different random seeds, and cross-validation, especially for DCA1 and XCAD.

Third, although we added an exploratory method-level statistical analysis based on the aggregated comparison results, this analysis has important limitations. The Wilcoxon signed-rank tests were performed across baseline methods using the reported mean values, rather than across individual test images. Similarly, the SD and 95% CI reported in Table 9 reflect variability across baseline-method-level improvements rather than repeated-run variability. Therefore, the results indicate that the proposed method achieved consistent positive improvements over the compared methods at the aggregated method level, but they do not replace rigorous image-level paired statistical testing. Because per-image paired scores or prediction masks for all competing methods were not available in the current comparison setting, we could not reliably estimate image-level variability, run-to-run variability, or confidence intervals across repeated training runs. Future work should include image-level paired comparisons, repeated training with different random seeds, and cross-validation, especially for smaller datasets such as DCA1 and XCAD.

Finally, the model introduces specialized modules and therefore remains more complex than plain encoder-decoder baselines. Recent studies have also emphasized the importance of structural feature enhancement, boundary-aware refinement, efficient architectural optimization, and feature fusion in medical image analysis [47, 48]. In addition, dynamic convolutional mechanisms have been explored in chest X-ray image captioning and radiology report understanding, suggesting the broader relevance of adaptive convolutional representations in radiological image analysis [49, 50]. In line with these directions, future work should explore pruning, quantization, knowledge distillation, and lightweight deployment strategies to improve applicability in real-time or resource-limited clinical environments. Temporal modeling across angiographic sequences may also further improve vessel continuity and reduce frame-level ambiguity.

Conclusions

In conclusion, the proposed vascular-aware segmentation framework provides a promising solution for X-ray angiography vessel segmentation. By jointly enhancing vascular structural modeling and texture recovery, it improves thin-vessel delineation, preserves topological continuity, and achieves favorable performance across three public datasets. Further validation using larger multi-center cohorts, image-level statistical testing, and clinically oriented evaluation is needed before translation into routine clinical use.

Abbreviations

Acc

Accuracy

ARCADE

Automatic Region-based Coronary Artery Disease Diagnostics using X-ray Angiography Images

BCE

Binary cross-entropy

CA

Channel attention

CNN

Convolutional neural network

CVAT

Computer Vision Annotation Tool

DCA1

Database of Coronary Artery Angiograms 1

Dice

Dice coefficient

DSnC

Dynamic snake convolution

FPS

Frames per second

IoU

Intersection over union

MoE

Mixture-of-experts

PGS

Parallel guidance strategy

SA

Spatial attention

TED

Texture-Enhanced Decoder

VA-MoE

Vascular-Aware Mixture-of-Experts

ViT

Vision Transformer

XCAD

X-ray angiography coronary artery disease dataset

XRA

X-ray angiography

Author contributions

Zhongjian Ju: Conceptualization, Methodology, Software, Formal analysis, Visualization, Writing - original draft.Qingnan Wu: Clinical investigation, Medical interpretation, Validation, Writing - review and editing.Wanqiang Cai: Methodology, Technical validation, Formal analysis, Writing - review and editing.Zidong Liu: Software, Data curation, Experimental support, Writing - review and editing.Ruigang Ge: Clinical resources, Medical interpretation, Writing - review and editing.Yimeng Li: Clinical consultation, Data interpretation, Writing - review and editing.Jinyuan Wang: Investigation, Data interpretation, Writing - review and editing.Zaichao Lu: Clinical consultation, Investigation, Writing - review and editing.Zeping Cui: Resources, Clinical coordination, Writing - review and editing.Jiasong Wu: Conceptualization, Supervision, Project administration, Funding acquisition, Writing - review and editing.Baolin Qu: Supervision, Clinical guidance, Writing - review and editing.Zhongjian Ju and Qingnan Wu contributed equally to this work. All authors read and approved the final manuscript.

Funding

This work was supported in part by the National Key Research and Development Program of China under Grant 2025YFE0150500, in part by the National Natural Science Foundation of China under Grant 62476055, in part by the innovation project of Jiangsu Province under Grants BZ2023042 and BY2022564, and in part by the Fundamental Research Funds for the Central Universities under Grant 2242024k30035.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This study used only publicly available anonymized X-ray angiography datasets and did not involve new recruitment, intervention, or identifiable personal information collected by the authors. Therefore, additional ethics approval and informed consent to participate were not required for this secondary analysis.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.