Vascular Morphology-Driven Augmentation Pipeline for Vessel Segmentation

Abstract

Accurate vessel segmentation is crucial for diagnosing cerebrovascular and retinal diseases, but the complexity of vessel structures makes manual annotation challenging, resulting in a lack of annotated data. Additionally, generated data often struggle to balance realism and diversity in vessel morphology. To address this, we provide a theoretical description of vessel structure and propose a vascular morphology-driven augmentation pipeline (VMDAP) for both feature reconstruction and vessel generation based on real vascular morphological characteristics. The pipeline consists of three components: the background restoration block (BRB), the vessel structure feature enhancement block (VSFEB), and the augmented data generation network (ADGN). VMDAP strategically separates vessel data into foreground and background, restores the background, reconstructs vessel masks features, and reintegrates them to generate new data. The ADGN incorporates vessel texture priors and employs a vessel detail enhancement loss to improve the realism of details. To address the performance characteristics of the VMDAP algorithm, this paper also introduces a novel loss function, DAclDice, which focuses on vessel branch structures to reduce false positives. Extensive experiments on six datasets across three representative network architectures validate the effectiveness of VMDAP and structure-driven loss function (DAclDice). Compared to nnU-Net with clDice loss, our method achieves an average Dice coefficient improvement of 2.15, a clDice improvement of 1.16, and a significant 36.17% reduction in 95th percentile Hausdorff distance (HD95). Excitingly, additional experiments demonstrate that VMDAP outperforms full real-data training when only half the real data are used.

Graphical abstract

Keywords

vessel segmentation / morphology-driven augmentation / synthetic vessel image generation / data augmentation / topology-aware loss / DAclDice

Introduction

Vessel segmentation holds paramount significance in clinical diagnosis, therapeutic interventions, and medical research. The precision in vessel segmentation outcomes substantially enhances the evaluation of vessel-related pathologies[1,2], providing crucial navigational cues for surgical interventions[3,4]. Compared to traditional vessel segmentation methods[5,6], data-driven deep learning approaches[7,8] offer significant improvements in segmentation accuracy and detail. However, the complexity of vessel structures makes manual annotation challenging, and privacy concerns limit access to vessel segmentation datasets, restricting the availability of labeled data.

To address the issue of limited labeled data, many methods focus on synthesizing vascular images. Some use features from labeled data[9,10], but the realism of these images is constrained by the generation algorithms. Other approaches enhance realism by generating blood vessels based on features from real images[11], but these methods lack diversity, generating only variations of the same mask. To improve diversity, some studies use manually designed algorithms to simulate realistic vessel distributions[12,13], but these methods suffer from high complexity and limited transferability.

To address these challenges, this paper proposes a vascular morphology-driven augmentation pipeline (VMDAP) and the corresponding DAclDice. The method aims to simultaneously extract both the background and vessel structures from real data, enhance vascular image features based on structural information, and generate new data, ultimately improving the segmentation accuracy of the algorithm. VMDAP consists of three main components: the background restoration block (BRB), the vessel structure feature enhancement block (VSFEB), and the augmented data generation network (ADGN). The process involves foreground separation, background restoration, vessel feature enhancement, and reintegration, resulting in new vessel images with distinct structural features and backgrounds. The first two modules are implemented using a combination of image processing techniques, while the ADGN leverages a vessel texture prior-guided transformer and a vessel detail enhancement loss to generate high-fidelity reconstructed vessel images. Extensive experiments are conducted to validate and summarize the performance characteristics of VMDAP, and based on its limitations, a novel loss function, DAclDice, is introduced to improve the network’s ability to capture the branch structures of vessels.

In summary, the contributions of this paper are as follows:

•We propose a VMDAP based on a theoretical description of vessel morphology. The pipeline consists of the BRB, VSFEB, and ADGN, which generate vascular data from background and reconstructed masks, ensuring both realism and diversity in vessel morphology.

•In ADGN, we employ a transformer architecture that integrates vessel texture priors and a vessel detail enhancement loss function to incorporate real texture information and enhance the network’s focus on vessel details.

•To address the limitations observed in VMDAP, we proposed the DAclDice loss function, which aims to improve the algorithm’s attention to vessel continuity while also emphasizing the width of vessel branches.

Traditional vessel segmentation methods typically rely on handcrafted designs tailored for specific data. Khan et al.[5] designed multiple filters for denoising, enhancement, and vessel segmentation in retinal images. Memari et al.[6] proposed an approach that combines contrast-limited adaptive histogram equalization and multiple filters for retinal image enhancement, followed by the use of a genetic algorithm-enhanced spatially fuzzy c-means method to extract the initial vessel segmentation map. Wu et al.[14] employed defogging, enhancement, and green channel extraction to preprocess blood vessel images in cranial windows. Nevertheless, these methods are manually designed for specific tasks, which limits their generalization capabilities. With the advancement of deep learning, the introduction of fully convolutional networks (FCN)[15] and U-Net[16] significantly improved segmentation precision. Wang et al.[17] applied U-Net to retinal vessel segmentation, surpassing previous methods. Li et al.[18] introduced a connection-sensitive Attention U-Net, improving the connectivity of segmentation results. Shit et al.[19] proposed the clDice loss function, based on topological information, to enhance the connectivity of segmented tubular structures. Additionally, some vessel segmentation studies focus on domain adaptation[20], unsupervised learning methods[21], and self-supervised learning algorithms[22].

To address the issue of limited labeled data for vessel segmentation, several approaches focus on generating synthetic blood vessel images. Some studies generate vascular images based on characteristics derived from annotated data. For example, Bonaldi et al.[10] utilized an active shape model (ASM) to synthesize reliable vessel shapes, and Pedro et al.[9] designed a generative adversarial network (GAN) that first generates vessel labels and then creates realistic vessels. Furthermore, these methods heavily depend on the algorithm’s design, and the generated vessels often suffer from breaks and limited realism, which hinders downstream segmentation performance on real data[13]. To improve the realism of vessel structures, some methods generate blood vessels using image features extracted from real vessels. For example, Zhang et al.[11] generated realistic vessel images by applying Sobel operators to extract vessel contours. However, these methods only generate different realizations of the same mask and do not account for background details, limiting both the diversity of the generated data and the realism of the background. To address the diversity issue in generated vessels, many studies have proposed manually designed vessel generation algorithms to simulate realistic vessel distributions. Menten et al.[12] designed vessel distributions based on physiology and applied image processing techniques for denoising and enhancement. Linus et al.[13] replaced the computationally expensive physiology-based simulation with a statistical angiogenesis model and used contrast-adaptive enhancement for images. While these approaches have produced large-scale datasets that perform well in tasks with specific vascular distributions, they suffer from high design complexity and limited transferability.

Therefore, to improve the segmentation performance of downstream tasks, vessel augmentation algorithms must balance both the realism and diversity of generated vessels. The data augmentation method proposed in this paper took into accounts the structural characteristics of vessels, constructed new masks based on real vascular masks, and combined realistic background and vascular textures to generate new vascular data, thus ensuring both the realism and diversity of the generated blood vessels.

Materials and Methods

VMDAP

VMDAP initiates from authentic vessel images, utilizing images and labels from the training set to achieve the separation, processing, and recombination of vessels and background, with the objective of augmenting the dataset. The following is the theoretical modeling of this problem.

Vessel images can be considered as a spatial composition of the vessel foreground and the background. Let I(x, y) represents the original vessel image, V(x, y) denotes the vessel foreground, and B(x, y) signifies the ideal background devoid of vessels. To rigorously define this composition, we introduce a binary vessel mask M(x, y)∈{0,1}. The mask is defined such that M(x, y) = 1 if the pixel at location (x, y) belongs to the vessel region (i.e., V(x, y) ≠ 0), and M(x, y) = 0 otherwise.

Using the Hadamard product (element-wise multiplication, denoted by ), the original vessel image I(x, y) can be mathematically formulated as:

\begin{array}{}\text{(1)}& \begin{array}{c}I\left(x,y\right)=B\left(x,y\right)\odot \left(1-M\left(x,y\right)\right)+V\left(x,y\right)\end{array}\end{array}

Consequently, the directly obtainable background B₀(x, y) from the training set, which contains zero-valued “holes” at the vessel locations, is rigorously derived as:

\begin{array}{}\text{(2)}& \begin{array}{c}{B}_0\left(x,y\right)=B\left(x,y\right)\odot \left(1-\text{M}\left(x,y\right)\right)\end{array}\end{array}

For the task of vessel data augmentation, utilizing B₀(x, y) as the background for subsequent composition is unfeasible due to the presence of original vessel masks. Hence, in proposed methodologies, the objective is to acquire a background (F(B₀(x, y))) closely approximating the ideal background B(x, y), expressed as:

\begin{array}{}\text{(3)}& \begin{array}{c}\overline{F\left(B_0\left(x,y\right)\right)}=\arg {\text{min}}_{\text{F}}\left|\left|F\left(B_0\left(x,y\right)\right)-B\left(x,y\right)\right|\right|\end{array}\end{array}

where \overline{F\left(B_0\left(x,y\right)\right)} signifies the ideal background extraction method. However, obtaining a vessel background without a mask in practical scenarios is unattainable. To overcome this limitation, we introduce the BRB, a module incorporating filtering and morphological processing to eliminate the vessel masks. This results in the generation of a vessel background B′(x, y) suitable for subsequent data synthesis:

\begin{array}{}\text{(4)}& \begin{array}{c}{B}^{\prime }\left(x,y\right)=\text{BRB}\left(B_0\left(x,y\right)\right)=\text{BRB}\left(I\left(x,y\right)-V\left(x,y\right)\right)\end{array}\end{array}

The algorithmic procedure of (BRB) is illustrated by the content in the left blue box of Figure 1. The operational mechanism consists of three sequential steps based on an adaptive kernel size (determined logarithmically by the image dimensions):

(1)Morphological pre-processing. The masked background first undergoes a median filter, followed by three successive morphological erosion operations, and another median filter. This process effectively shrinks the blank masked regions and expands the valid background boundaries.

(2)Iterative local inpainting (PNNA). To fill the remaining missing pixels (identified by high-intensity thresholds, e.g., > 240), a priority nearest neighbor assignment (PNNA) algorithm is applied. For each missing pixel, the algorithm searches a local bounding box (e.g., a 19 × 19 local window) and assigns the color value of the first encountered valid background pixel. This operation is iterated multiple times (e.g., 3 iterations) to progressively fill larger structural gaps.

(3)Post-processing smoothing. Finally, to eliminate any blocky artifacts introduced during the local inpainting process and restore a natural, seamless background texture, the image is refined using a combination of median filtering and average blurring.

After acquiring the generated background B′(x, y) for each frame in the training set using BRB and obtaining the vessel masks V(x, y) from labels, this study emphasizes the generation of new vessels based on the inherent characteristics of blood vessels. The precision of actual vessel segmentation tasks is frequently hindered by the intricate structures at the terminal ends of vessels. Therefore, the diversity of terminal structures becomes a pivotal consideration in the process of generating new vessel masks.

Consider a vessel composed of three segments, represented as:

\begin{array}{}\text{(5)}& \begin{array}{c}V\left(x,y\right)=Tr\left(x,y\right)+Br\left(x,y\right)+Te\left(x,y\right)\end{array}\end{array}

where Tr(x, y) denotes the main trunk of the vessel, Br(x, y) represents its branches, and Te(x, y) signifies the terminals of the vessel. Ideally, the thickness of these three components gradually diminishes. This method employs an erosion operation on the original vessel to derive a generated vessel with novel terminal features, expressed as:

\begin{array}{}\text{(6)}& \begin{array}{c}{V}^{\prime }\left(x,y\right)=\text{Erode}\left(V\left(x,y\right)\right)\end{array}\end{array}

Due to variations in thickness, En(x, y) is initially set to 0 through erosion. Consequently, the new vessel can be succinctly represented as:

\begin{array}{}\text{(7)}& \begin{array}{c}{V}^{\prime }\left(x,y\right)=T{r}^{\prime }\left(x,y\right)+T{e}^{\prime }\left(x,y\right)\end{array}\end{array}

In the given equation, when Te(x, y) is set to 0 through erosion, the original vessel’s branches Br(x, y) undergo transformation into the terminals of the new vessel. Consequently, the generation of the newly formed vessel V′(x, y) is achieved. The proposed VSFEB encompasses not only the aforementioned operations but also incorporates fundamental operations such as rotations and adjustments to the brightness of the new background and the new blood vessels. These operations contribute to the generation of the enhanced vessel image V′ and the modified background B′. The algorithmic workflow of VSFEB is delineated in the right red box of Figure 1. The entire amalgamation process of vessel masks and background can be concisely expressed as:

\begin{array}{}\text{(8)}& \begin{array}{c}{I}^{\prime }\left(x,y\right)=\text{VSFEB}\left(B^{\prime }\left(x,y\right),V\left(x,y\right)\right)\end{array}\end{array}

Subsequently, the ADGN generates vessel data based on above mask images. The comprehensive workflow of VMDAP can be articulated as follows:

\begin{array}{}\text{(9)}& \begin{array}{c}\text{VMDAP}\left(I,V\right)=\text{ADGN}\left(\text{VSFEB}\left(\text{BRB}\left(I,V\right),V\right)\right)\end{array}\end{array}

In this representation, I and V correspond to the vessel image and the vessel masks derived from the ground truth labels, respectively.

Augmented data generation network

The ADGN is employed to generate vessel images from vessel masks with backgrounds. As shown in the lower purple box of Figure 1, the proposed generation network is based on the Transformer architecture, with the core structure derived from Restormer[23]. It includes downsampling Transformer blocks, upsampling Transformer blocks, and a refinement layer. For the vessel generation task, the network integrates vessel texture priors through explicit loss supervision during the training phase. Specifically, the pure vessel foreground is first extracted using the ground-truth segmentation mask to eliminate background interference. The Sobel operator is then applied to this isolated region to obtain a gradient map, which serves as the vessel texture prior. During training, the network computes a texture loss between the Sobel gradient of the generated vessels and this authentic prior. This forces the generator to focus heavily on synthesizing realistic, high-frequency vascular textures.

In real vessel images, regions of high brightness often occur due to lighting or contrast agents, leading to significant increases in vessel brightness. Additionally, the vessel wall, distinct from the lumen, casts shadows due to its larger light occlusion. These two characteristics—bright highlights and shadowed vessel walls—are crucial details in vessel images. To improve the network’s sensitivity to these details, we proposed a vessel detail enhancement loss, which implicitly acts as a dynamic spatial weighting map. In normalized image space, the shadowed vessel walls and bright highlights correspond to extreme pixel intensities (approaching 0 and 1, respectively). To amplify the learning focus on these areas, we designed a spatial weight mapping based on the parabolic function of pixel intensities x: W(x) = 1–x(1–x). this function assigns maximal weights to extreme values (highlights and shadows) while marginally suppressing the mid-tone regions. By applying this spatial weight W(x) element-wise to both the generated predictions and ground-truth targets prior to calculating the L₁ loss, the network is strictly constrained to accurately reconstruct the high-frequency micro-structures and contrast agents. The detailed algorithmic implementation is further provided in Supplementary Figure S1. For the experiments on each dataset, the training data were cropped to 128 × 128 and used for training the ADGN.

DAclDice

The centerline dice loss (clDice) takes into consideration the topological characteristics and connectivity of blood vessel and quantifies the connectivity loss by assessing the alignment of centerlines between the segmented image and the ground truth labels. Referencing Supplementary Figure S1 in practical scenarios, the algorithm demonstrates heightened focus on vessel terminals due to their narrower width compared to vessel branches and main trunks, resulting in a more substantial presence of centerlines in the structure. This heightened sensitivity to fine structures increases the algorithm’s susceptibility to false positives induced by background influences.

Building upon the insights gained from the analysis of clDice, this research introduces DAclDice. As depicted in Supplementary Figure S1, DAclDice not only engages in a comparative analysis with the centerline but also extracts morphological features of the branches connected to vessel terminals, integrating them into the loss function. The proposed loss aims to amplify the responsiveness of algorithm to the branches associated with terminals, with the objective of isolating isolated points and mitigating false positives. The algorithmic workflow for extracting these branches is depicted in Supplementary Figure S2, where the symbol in the table denotes element-wise multiplication.

After obtaining the branch information from the predicted results and ground truth values, DAclDice can be calculated using the following formulas:

\begin{array}{}\text{(10)}& \begin{array}{c}{T}_p\left(B_P,V_L\right)=\frac{\left|B_P\cap V_L\right|}{\left|B_P\right|}\end{array}\end{array}

\begin{array}{}\text{(11)}& \begin{array}{c}{T}_s\left(B_L,V_P\right)=\frac{\left|B_L\cap V_P\right|}{\left|B_L\right|}\end{array}\end{array}

\begin{array}{}\text{(12)}& \begin{array}{c}\text{DA}\left(V_P,V_L\right)=2\frac{T_p\left(B_P,V_L\right)\cdot T_s\left(B_L,V_P\right)}{T_p\left(B_P,V_L\right)+T_s\left(B_L,V_P\right)}\end{array}\end{array}

where DA represents DAclDice V_P and V_L are the predicted segmentation mask and ground truth mask, respectively. B_P and B_L represent the extracted branch information from V_P and V_L. T_p(B_P,V_L) stands for topological accuracy, and T_s(B_L,V_P) represents the corresponding sensitivity. The final soft-DAclDice loss is a combination of the Dice function, clDice function, and DAclDice function, expressed as:

\begin{array}{}\text{(13)}& \begin{array}{c}\text{L}=\left(1-\text{α}-\text{β}\right)\left(1-\text{Dice}\right)+\alpha \left(1-\text{clDice}\right)+\beta \left(1-\text{DAclDice}\right)\end{array}\end{array}

where α and β are scaling factors.

Results

Experimental design

To validate the effectiveness of the proposed algorithm, three sets of experiments were meticulously designed. The first set evaluated the performance of VMDAP across five datasets and three representative algorithms with different architectures. The second set used nnU-Net and clDice as baselines to assess the combined performance of VMDAP and DAclDice, as well as their individual performances across the five datasets. Additionally, we assessed VMDAP’s performance under limited real data conditions. The third set aimed to investigate the contribution of the two components of the ADGN to the improvement of downstream segmentation accuracy. By selecting these state-of-the-art and highly representative previously reported models, we aimed to demonstrate that VMDAP and DAclDice can serve as universally applicable enhancements rather than competing as a standalone network.

The experiments were conducted on five distinct retinal vessel segmentation datasets and a intracranial artery segmentation dataset: DRIVE[24], Chasedb[25], HRF-Healthy, HRF-Glaucomatous, HRF-Diabetic[26], and DIAS[27]. These datasets encompass a variety of imaging devices, resolutions, and contain vessel images depicting both normal and pathological conditions.

To cover different network architectures, three representative algorithms were selected: nnU-Net[28] for CNN-based approaches, TransU-Net[29] for Transformer-based methods, and U-Mamba[30] for networks based on the Mamba architecture.

To comprehensively evaluate the distinct characteristics of each method, three performance metrics were employed. The Dice[31] coefficient measured the overlap between the predicted segmentation and the ground truth labels. The clDice metric assessed the connectivity of extracted vessels and the structural elements in the image. HD95 was used to evaluate the maximum distance between two points sets, providing a detailed assessment of segmentation performance, particularly in complex regions such as edges.

Experimental configuration

The experimental setup utilized the computational power of an NVIDIA GeForce RTX 3090 GPU with 24GB of memory, while the U-Mamba-related experiments were conducted on an NVIDIA GeForce RTX 4090 GPU. All implementations were carried out on the PyTorch[32] public platform.

For each dataset, the publicly available data were evenly split into training and testing sets based on their sequence numbers. In the experiments assessing the performance of the VMDAP algorithm, VMDAP applied two rounds of data augmentation to each background image in the training set. After augmentation, the number of samples in the training set was tripled compared to the original dataset. To ensure fairness in training, all comparative experiments within the same group used identical iterations and learning rates. This mitigated the influence of sample size differences across datasets.

For different algorithms, the default configurations provided by the official specifications were used. For TransU-Net, the input image size was set to 224 × 224, with a batch size of 8. The initial learning rate was 0.01, and the learning rate was updated using the Poly method with a power value of 0.9. For nnU-Net and U-Mamba, default settings, including adaptive input size and batch size, were applied, and the learning rate update strategy was the same as that for TransU-Net.

Performance evaluation of VMDAP

In this experimental paradigm, we conducted a comprehensive assessment of the VMDAP algorithm, leveraging the TransU-Net and nnU-Net methodologies across five distinct datasets. The quantitative outcomes of these experiments are meticulously documented in Table 1.

VMDAP demonstrates consistent performance improvements across models. TransU-Net shows average gains of 0.64 in Dice and 1.70 in clDice, with a 13.33% reduction in HD95. Similarly, nnU-Net improved by 0.74 in Dice and 1.00 in clDice, reducing HD95 by 25.88%. U-Mamba achieved gains of 0.61 in Dice, 0.88 in clDice, and 12.12% lower HD95. These quantitative improvements confirmed VMDAP’s effectiveness.

Visual analysis confirms that VMDAP-enhanced models better capture fine details and maintain vessel continuity while reducing false positives. However, challenges remain in eliminating false positives and ensuring complete vessel continuity, partly due to increased sensitivity to microscopic terminal structures introduced during generation.

To address these issues, this paper employs soft-clDice and introduces the optimized loss function soft-DAclDice to improve false positives and vessel continuity. Experimental results are shown in the next section.

Combined performance experiment

We evaluated the proposed algorithms—VMDAP and DAclDice—against baseline methods (nnU-Net and soft-clDice) across five datasets. For the clDice loss group, Dice and clDice weights were set to 0.5 each, following the original clDice paper. In the DAclDice group, clDice weight was 0.4, with an additional 0.1 for DAclDice. Detailed results are in Table 2.

While traditional multi-run statistical significance testing was constrained by the substantial computational overhead of training heavy architectures (e.g., nnU-Net, U-Mamba) across six diverse datasets, the robustness of our method is strongly evidenced by its consistent behavior. Quantitative results show significant improvements: VMDAP achieved a +1.23 increase in Dice, +1.11 in clDice, and a 24.05% reduction in HD95 compared to the baseline. Similarly, soft-DAclDice improved by +1.12 in Dice, +0.47 in clDice, and −20.07% in HD95. The combined method outperformed both, with gains of +2.15 in Dice, +1.16 in clDice, and −36.17% in HD95, demonstrating strong generalization across all datasets. This extensive cross-domain consistency demonstrates that the performance gains are highly stable and generalizable, rather than the result of random variance. Visual results are shown in Figure 2.

Qualitative analysis revealed that both VMDAP and soft-DAclDice reduced false positives, with soft-DAclDice improving vessel continuity. The combined method further enhanced false positive suppression and refined vessel details. However, some prediction gaps remain (Figure 2, last column), even where soft-DAclDice corrected discontinuities (blue arrows).

Performance of VMDAP with limited data

To further evaluate the performance of VMDAP, we conducted experiments with limited data on high-resolution datasets HRF-H, HRF-G, and HRF-D. Specifically, we trained ADGN with a reduced amount of real data and generated new data accordingly, then compared the performance of this dataset with results obtained using a larger amount of real data. In this experimental setup, our proposed method used four frames of real data and four frames of augmented data generated from real data, while the comparison group used eight frames of real data as in previous experiments. The results, shown in Table 3, are promising: our method, with only half the real data, outperforms the fully real-data-based method across most metrics. This demonstrates the strong performance of VMDAP and its enhanced capability to improve performance with limited data. It is worth noting that while the relative performance gain is most pronounced in the limited data regime, VMDAP continues to provide strictly positive and significant improvements even when applied to the full dataset (as previously demonstrated in Table 1). Furthermore, because VMDAP is strictly bounded by real anatomical masks and texture priors, it avoids the out-of-distribution risks typically associated with over-augmentation.

Ablation study of ADGN

An ablation study was conducted on two components of ADGN within the VMDAP framework. For the quantitative evaluation, the effectiveness of the generated data was assessed based on the final metrics of the vessel segmentation downstream task. The ablation experiments were carried out on the Chasedb dataset, where the effects of excluding vessel texture priors and the absence of the vessel detail enhancement loss were evaluated. The results are summarized in Table 4. Specifically, when texture priors are excluded, the generative network tends to produce flat, homogeneous tubular structures that lack the natural central light reflex characteristic of real retinal vessels. Furthermore, without the Vessel Detail Enhancement loss, the generated vessels suffer from blurred boundary walls and washed-out highlights, making them appear artificially overlaid rather than naturally integrated into the background lighting. With the incorporation of both modules, the ADGN synthesizes vessels with highly realistic specular highlights and sharp, accurate shadow gradients. This extreme fidelity in texture and microscopic detail preservation directly translates to the enhanced downstream segmentation accuracy of fine capillaries, as previously visualized in Figure 2.

Conclusion

This study presents an innovative data augmentation method, VMDAP, specifically designed for vessel segmentation, complemented by a novel loss function. VMDAP constructs new vascular masks using a structure-guided augmentation approach based on real background and vessel ground truth masks, and designs a generative network to create new data, ensuring both the authenticity and diversity of the generated vascular structures. DAclDice, proposed in conjunction with VMDAP, enhances the network’s loss calculation for vascular branches through a structure-guided approach to address the false positive issues left by VMDAP. Extensive experiments validate the effectiveness of both individual methods and their combination, with ablation studies confirming the importance of the ADGN component. Overall, this study introduces a vascular enhancement approach that emphasizes vascular structural features and structural authenticity, which not only advances vessel segmentation performance but also provides a new perspective for related fields. Furthermore, while the current extensive evaluations are focused on two-dimensional (2D) retinal and intracranial DSA images, the core theoretical framework of VMDAP—leveraging universal anatomical topologies such as tree-like branching and tubular structures—holds significant potential for generalization to other clinical modalities. Future work will focus on adapting this pipeline for three-dimensional (3D) modalities like computed tomography (CT) and magnetic resonance (MR) angiography by extending the morphological and generative modules into the volumetric domain. Additionally, generalizing VMDAP to Ultrasound imaging will be explored by incorporating modality-specific physical priors to account for acoustic shadowing and speckle noise, thereby broadening the clinical impact of morphology-driven augmentation.

References

[1] Hunter JM, Kwan J, Malek-Ahmadi M, et al. Morphological and pathological evolution of the brain microcirculation in aging and Alzheimer’s disease. PLoS One. 2012;7(5):e36893.

[2] Joutel A, Monet-Leprêtre M, Gosele C, et al. Cerebrovascular dysfunction and microcirculation rarefaction precede white matter lesions in a mouse genetic model of cerebral ischemic small vessel disease. J Clin Invest. 2010;120(2):433-445.

[3] Selle D, Preim B, Schenk A, Peitgen HO. Analysis of vasculature for liver surgical planning. IEEE Trans Med Imaging. 2002;21(11):1344-1357.

[4] Dolati P, Eichberg D, Golby A, Zamani A, Laws E. Multimodal navigation in endoscopic transsphenoidal resection of pituitary tumors using image-based vascular and cranial nerve segmentation: a prospective validation study. World Neurosurg. 2016;95:406-413.

[5] Khan KB, Siddique MS, Ahmad M, Mazzara M. A hybrid unsupervised approach for retinal vessel segmentation. Biomed Res Int. 2020;2020:8365783.

[6] Memari N, Ramli AR, Sarip MIB, Mashohor S, Moghbel M. Retinal blood vessel segmentation by using matched filtering and fuzzy c-means clustering with integrated level set method for diabetic retinopathy assessment. J Med Biol Eng. 2019;39:713-731.

[7] Meyer A, Mehrtash A, Rak M, et al. Domain adaptation for segmentation of critical structures for prostate cancer therapy. Sci Rep. 2021;11(1):11480.

[8] Zhi Y, Hau WK, Zhang H, Gao Z. Vessel contour detection in intracoronary images via bilateral cross-domain adaptation. IEEE J Biomed Health Inform. 2023;27(7):3314-3325.

[9] Costa P, Galdran A, Meyer MI, et al. End-to-End adversarial retinal image synthesis. IEEE Trans Med Imaging. 2018;37(3):781-791.

[10] Bonaldi L, Menti E, Ballerini L, Ruggeri A, Trucco E. Automatic generation of synthetic retinal fundus images: vascular network. Procedia Comput Sci. 2016;90:54-60.

[11] Zhang T, Fu H, Zhao Y, et al. SkrGAN: sketching-rendering unconditional generative adversarial networks for medical image synthesis. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2019. Lecture Notes in Computer Science. Vol 11767. 2019:777-785.

[12] Menten MJ, Paetzold JC, Dima A, et al. Physiology-based simulation of the retinal vasculature enables annotation-free segmentation of OCT angiographs. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2022. Lecture Notes in Computer Science. Vol 13431. 2022:330-340.

[13] Kreitner L, Paetzold JC, Rauch N, et al. Synthetic optical coherence tomography angiographs for detailed retinal vessel segmentation without human annotations. IEEE Trans Med Imaging. 2024;43(6):2061-2073.

[14] Wu Y, Oda M, Hayashi Y, et al. Blood vessel segmentation from low-contrast and wide-field optical microscopic images of cranial window by attention-gate-based network[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022:1864-1873.

[15] Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2015:3434-3440.

[16] Ronneberger O, Fischer P, Brox T. U-net: convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer Assisted Intervention- MICCAI 2015. Lecture Notes in Computer Science. Vol 9351. 2015:234-241.

[17] Das S, Chakraborty S, Majumder S. Assessment of retinal blood vessel segmentation using U-Net model: a deep learning approach. Franklin Open. 2024:article ID 100143.

[18] Li R, Li M, Li J, Zhou Y. Connection sensitive attention U-Net for accurate retinal vessel segmentation. arXiv. Preprint posted online April 23, 2019; arXiv:1903.05558v2.

[19] Shit S, Paetzold JC, Sekuboyina A, et al. clDice-a novel topology-preserving loss function for tubular structure segmentation. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2021:16560-16569.

[20] Bermúdez-Chacón R, Márquez-Neila P, Salzmann M, Fua P. A domain-adaptive two-stream U-Net for electron microscopy image segmentation. Paper presented at: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI); April 4-7, 2018; Washington, DC, USA.

[21] Hoover A, Kouznetsova V, Goldbaum M. Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response. IEEE Trans Med Imaging. 2000;19(3):203-210.

[22] Ma Y, Hua Y, Deng H, et al. Self-supervised vessel segmentation via adversarial learning. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE; 2021:7516-7525.

[23] Zamir SW, Arora A, Khan S, Hayat M, Khan FS, Yang MH. Restormer: efficient transformer for high-resolution image restoration. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE; 2022:5718-5729.

[24] Staal J, Abràmoff MD, Niemeijer M, Viergever MA, van Ginneken B. Ridge-based vessel segmentation in color images of the retina. IEEE Trans Med Imaging. 2004;23(4):501-509.

[25] Owen CG, Rudnicka AR, Nightingale CM, et al. Retinal arteriolar tortuosity and cardiovascular risk factors in a multi-ethnic population study of 10-year-old children; the child heart and health study in england (CHASE). Arterioscler Thromb Vasc Biol. 2011;31(8):1933-1938.

[26] Budai A, Bock R, Maier A, Hornegger J, Michelson G. Robust vessel segmentation in fundus images. Int J Biomed Imaging. 2013;2013:154860.

[27] Liu W, Tian T, Wang L, et al. DIAS: A dataset and benchmark for intracranial artery segmentation in DSA sequences. Med Image Anal. 2024;97:103247.

[28] Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18(2):203-211.

[29] Chen J, Lu Y, Yu Q, et al. TransUNet: transformers make strong encoders for medical image segmentation. arXiv. Preprint posted online February 8, 2021; arXiv: 2102.04306v1.

[30] Ma J, Li F, Wang B. U-Mamba: enhancing long-range dependency for biomedical image segmentation. arXiv. Preprint posted online January 9, 2024; arXiv: 2401.04722v1.

[31] Milletari F, Navab N, Ahmadi SA. V-net: fully convolutional neural networks for volumetric medical image segmentation. In: 2016 Fourth International Conference on 3D Vision (3DV). IEEE; 2016:565-571.

[32] Paszke A, Gross S, Massa F, et al. PyTorch: an imperative style, high-performance deep learning library. arXiv. Preprint posted online December 3, 2019; arXiv:1912.01703v1.

RIGHTS & PERMISSIONS

The Author(s) 2026. This article is published by Higher Education Press at journal.hep.com.cn.