MoE-based Feature Adapter for Prompt-free Binary Coronary Artery Segmentation in X-ray Angiography
Proposed MoE-based feature adapter for prompt-free binary coronary artery segmentation, significantly improving cross-dataset generalization.
Key Findings
Methodology
This approach leverages parameter-efficient Vision Transformer adapters integrated with multiple lightweight experts. Using input-dependent top-k routing, the model dynamically selects relevant experts to refine vessel features at multiple transformer stages. The architecture embeds MoE modules within the encoder, enabling adaptive feature enhancement tailored to heterogeneous angiographic appearances. The training combines Dice and BCE losses to optimize vessel foreground detection. Experiments on MOSXAV demonstrate superior performance over U-Net, Attention U-Net, and nnU-Net, with significant improvements in Dice and IoU scores, and enhanced cross-dataset robustness validated on XACV.
Key Results
- On MOSXAV test set, the proposed method achieved a Dice score of 53.47%, IoU of 39.09%, outperforming baseline models, especially in capturing fine vessel details and suppressing background noise.
- External evaluation on XACV yielded Dice of 70.25% and IoU of 54.95%, indicating strong generalization capabilities across different imaging distributions.
- Ablation studies confirmed that multi-expert MoE modules with top-k routing provided consistent performance gains (~1.5% Dice increase) with limited parameter overhead, validating the effectiveness of the design.
Significance
This work addresses the longstanding challenge of robust coronary artery segmentation amid heterogeneous vessel appearances and complex backgrounds. By integrating a MoE mechanism, the model adapts to diverse angiographic patterns, improving accuracy and robustness crucial for clinical applications such as diagnosis, treatment planning, and image-guided interventions. The demonstrated cross-dataset performance suggests potential for broad deployment in real-world scenarios, reducing reliance on dataset-specific tuning.
Technical Contribution
The key innovation lies in embedding multiple lightweight experts within a Vision Transformer framework, combined with an input-dependent sparse routing strategy. This enables dynamic, context-aware feature refinement, surpassing traditional single-adapter approaches. The architecture balances parameter efficiency with high representational capacity, opening new avenues for scalable, adaptive medical image segmentation models. The method also provides a flexible template for integrating MoE modules into other dense prediction tasks.
Novelty
This is the first application of a MoE-based dynamic expert selection mechanism within a Transformer architecture for coronary artery segmentation. Unlike prior works relying on fixed feature transformations, the proposed model adaptively combines multiple specialized experts based on input appearance, effectively handling the heterogeneity of angiographic images. This innovation significantly advances the state-of-the-art in adaptive medical image segmentation.
Limitations
- Despite improved robustness, the model may still struggle with extremely low contrast or severe motion artifacts, leading to missed vessel segments.
- Increased model complexity due to multiple experts and routing mechanisms results in higher computational costs, which could hinder real-time deployment.
- Current evaluation focuses on static frames; incorporating temporal information could further enhance segmentation stability and accuracy in videos.
Future Work
Future research will explore integrating temporal modeling to leverage motion cues in angiography videos, aiming to improve continuity and measurement accuracy. Additionally, optimizing routing algorithms and model pruning could reduce inference costs, facilitating real-time clinical deployment. Extending the framework to multi-modal data and 3D volumetric segmentation also presents promising directions for comprehensive vascular analysis.
AI Executive Summary
Accurate segmentation of coronary arteries in X-ray angiography remains a critical challenge due to the vessels’ thin, tortuous, and often discontinuous nature, compounded by low contrast and complex backgrounds. Existing deep learning models like U-Net and Transformer-based architectures have made significant progress but still face limitations in handling heterogeneous angiographic appearances. To address this, the present study introduces a novel Mixture-of-Experts (MoE) feature adapter integrated into a Vision Transformer framework, designed to enhance the model’s adaptability and robustness.
The core innovation involves embedding multiple lightweight expert modules within the encoder’s transformer layers, coupled with an input-dependent top-k routing mechanism. This setup allows the model to dynamically select and fuse the most relevant experts for each feature, effectively refining vessel representations amidst diverse imaging conditions. The training combines Dice and BCE losses, ensuring precise vessel foreground segmentation.
Experimental results on the MOSXAV dataset demonstrate that the proposed method surpasses several state-of-the-art baselines, achieving a Dice score of 53.47% and IoU of 39.09%. Notably, in external validation on the XACV dataset, the model maintains strong performance with a Dice of 70.25% and IoU of 54.95%, evidencing excellent cross-dataset generalization. Ablation studies confirm that the multi-expert MoE structure and sparse routing significantly contribute to performance gains without excessive parameter increase.
This approach marks a significant step forward in medical image segmentation, offering a flexible, adaptive framework capable of handling the heterogeneity inherent in angiographic images. It opens avenues for more robust, scalable, and clinically applicable vascular analysis tools. Future work will focus on incorporating temporal dynamics and multi-modal data to further improve accuracy and clinical utility, ultimately aiding early diagnosis and intervention in coronary artery disease.
Deep Analysis
Background
血管分割作为医学影像分析中的关键任务，经历了从传统滤波、边缘检测到深度学习模型的快速演进。早期方法如Hessian滤波和血管增强滤波在复杂背景下表现有限。近年来，U-Net及其变体在多尺度特征提取方面取得突破，显著提升血管检测精度。Transformer模型引入后，长距离依赖建模增强，但在血管细节丰富、背景复杂的场景中仍存在鲁棒性不足的问题。多模态融合和半监督策略虽有所突破，但血管多样性和背景干扰仍未得到根本解决。血管的弯曲、断续和低对比特性使得自动化分割成为难题，亟需更具适应性和鲁棒性的模型架构。
Core Problem
核心问题在于血管的细长、弯曲和断续特性，加之低对比和背景干扰，导致单一模型难以同时兼顾血管细节和背景抑制。传统方法在复杂血管结构和异质背景中表现不稳定，容易漏检或误检。现有模型缺乏针对血管多样性和背景复杂性的自适应调节能力，限制了其临床应用推广。如何设计一种能根据局部特征动态调整的模型，提升血管检测的准确性和鲁棒性，是亟待解决的关键问题。
Innovation
本研究的创新在于引入多专家MoE机制，结合参数高效的AdaptFormer风格适配器，实现输入依赖的动态专家选择。具体创新包括：
- �� 多专家结构：多个轻量级适配器协同作用，丰富特征表达。
- �� 稀疏top-k路由：只激活最相关的专家，降低计算成本，提升效率。
- �� 多层次插入：在不同Transformer层中引入MoE模块，捕获不同尺度血管信息。
- �� 端到端训练：结合Dice和BCE损失，优化血管区域识别能力。这些创新共同增强模型对血管多样性和背景复杂性的适应性。
Methodology
- �� 输入：单帧血管造影图像，经过预处理和归一化。
- �� 编码器：基于Vision Transformer提取多层次特征。
- �� MoE适配器：在Transformer块中嵌入多个轻量级专家模块，每个专家通过瓶颈结构实现特征变换。
- �� 路由机制：利用池化特征，通过线性投影预测每个专家的重要性权重，采用top-k稀疏路由选择最相关的专家。
- �� 特征融合：激活的专家输出加权融合，生成自适应残差，细化血管特征。
- �� 多层次插入：在不同Transformer层中引入MoE适配器，捕获不同尺度血管信息。
- �� 解码器：融合多尺度特征，输出二值血管掩码。
- �� 损失函数：结合Dice和BCE，优化血管检测的准确性。
Experiments
采用MOSXAV数据集作为主要训练和测试数据，模型在验证集和测试集上进行性能评估。对比U-Net、Attention U-Net、nnU-Net等主流模型，指标包括Dice、IoU、Precision和Recall。超参数设置：专家数M=4，激活top-k=2，训练40轮，学习率3e-4。在XACV数据集上进行外部验证，测试模型跨域适应性。通过消融实验验证多专家MoE结构和稀疏路由的效果，分析参数增长与性能提升关系。
Results
在MOSXAV测试集，提出方法的Dice达53.47%，IoU为39.09%，优于所有对比模型，特别在血管细节恢复和背景抑制方面表现优异。在XACV外部验证中，Dice达70.25%，IoU为54.95%，展现出强大的跨域适应能力。消融实验显示，多专家MoE结构和top-k路由显著提升性能，参数增长有限，验证了设计的有效性。这些结果充分证明了模型在复杂血管结构和异质背景中的鲁棒性。
Plain Language Accessible to non-experts
想象你在厨房里做菜，锅里有许多不同的厨具和调料，每次做不同菜肴时都需要选择合适的工具和调料。比如炒菜时，有的菜需要多放盐，有的则要少放，甚至还要用不同的锅具。这个研究就像是给电脑配备了多个“厨具”，它可以根据不同的血管图片自动选择最合适的“厨具”来处理。这样，无论血管多细、多弯，背景多复杂，电脑都能灵活应对，把血管清楚地画出来。比起用一种方法，多个“厨具”让它更聪明、更灵活，也能更准确地帮医生找到血管，帮助诊断心脏病。就像厨师有了多把厨具，做菜就更得心应手，菜也会更好吃！
ELI14 Explained like you're 14
你喜欢做饭吗？想象一下，你有很多不同的厨具，比如炒锅、蒸锅、调料瓶，每次做菜都要根据菜的不同选择合适的工具和调料。这个研究就像是给电脑装上了很多“厨具”，它可以根据不同的血管图片自动挑选最合适的“厨具”来帮你画出血管。这样，无论血管长得多细、多弯，或者背景多复杂，电脑都能聪明地找到血管，把它们清楚地画出来。比起以前只用一种方法，这个新方法就像有了多把厨具，做菜更灵活，效果也更棒。它能帮医生更快更准地找到血管，早发现心脏病问题。就像厨师用多种厨具做出美味佳肴一样，这个技术让血管检测变得更聪明、更可靠！
Abstract
Accurate segmentation of coronary arteries in X-ray angiography videos is essential for quantitative coronary analysis and image-guided interventions. However, accurate segmentation remains challenging because coronary vessels are thin and exhibit low contrast, while the presence of catheters, guidewires, and complex anatomical background structures can further interfere with vessel delineation. Existing U-Net- and Transformer-based models provide strong baselines, but their shared feature-adaptation pathways may be insufficient for heterogeneous angiographic appearances. In this paper, we propose a prompt-free mixture-of-experts (MoE) feature adapter for binary coronary artery segmentation. Built upon parameter-efficient Vision Transformer adapters, the proposed method uses multiple lightweight experts with input-dependent top-$k$ routing to adaptively refine vessel-related features while limiting active computational cost. Experiments on MOSXAV and external evaluation on XACV show that the proposed method outperforms representative baselines and improves cross-dataset generalisation. These results suggest that MoE-based adapter learning is effective for robust coronary artery segmentation in X-ray angiography videos.