Source-linked AI summary
MoE-based Feature Adapter for Prompt-free Binary Coronary Artery Segmentation in X-ray Angiography
Lin Xi, Yingliang Ma
TL;DR
Coronary artery segmentation in X-ray angiography is challenged by thin, low-contrast vessels and heterogeneous interference. The paper proposes a prompt-free MoE feature adapter with sparse input-dependent routing, improving segmentation performance and cross-dataset generalisation on MOSXAV and XACV.
Problem
Single shared adapter pathways may inadequately handle heterogeneous angiographic appearances, limiting robust coronary artery segmentation needed for quantitative analysis and image-guided interventions.
Method
A prompt-free feature adapter uses multiple lightweight experts and input-dependent sparse top-k routing to adaptively refine features for binary coronary artery segmentation.
Results
Experiments on MOSXAV and external testing on XACV show improved segmentation performance and stronger cross-dataset generalisation than representative baselines.
Takeaways & Limitations
MoE-based adapter learning is a promising strategy for robust coronary artery segmentation in X-ray angiography videos.
Abstract
from arXiv · showhide
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.
1 Introduction
Automatic coronary artery segmentation in X-ray angiography is important for vessel analysis and clinical assistance but remains difficult because vessels are thin, tortuous, low-contrast, and unevenly visible. The paper addresses this limitation with a prompt-free MoE feature adapter using lightweight experts and sparse input-dependent routing, evaluated on MOSXAV and XACV.
- Motivation: Coronary artery segmentation supports quantitative vessel analysis, treatment planning, image-guided interventions, and downstream computer-assisted diagnosis.X-ray coronary angiography provides real-time coronary-artery visualisation during clinical procedures.
- Challenge: Thin, tortuous, curvilinear arteries with low contrast, uneven contrast-agent filling, and weak distal visibility make robust segmentation challenging.These characteristics can make vessels appear partially discontinuous in angiography videos.
- Related Work: Existing approaches include encoder-decoder networks such as U-Net, Attention U-Net, and nnU-Net, alongside newer Transformer-based segmentation models.These methods respectively provide hierarchical feature extraction with spatial detail recovery, self-configuring adaptation, or long-range dependency modelling.
- Proposed Method: The proposed prompt-free framework replaces a single AdaptFormer-style adapter with multiple lightweight experts and input-dependent routing for adaptive feature refinement.Parameter-efficient adapters offer compact trainable branches, while the proposed design dynamically combines relevant experts for each feature representation.
- Contributions: Sparse top-k expert routing improves adaptation flexibility while limiting active computational cost, and evaluation on MOSXAV and XACV demonstrates improved segmentation performance and cross-dataset generalisation over representative baselines.The framework is designed for binary coronary artery segmentation in X-ray angiography videos.
2 Method
The proposed method is an encoder-decoder segmentation network with transformer-based encoding and MoE feature adapters that adaptively refine intermediate features for binary coronary artery mask prediction. Multiple lightweight experts use input-dependent top-k routing to provide efficient, complementary refinements across transformer stages and heterogeneous angiographic appearances.
- Network formulation: The model predicts a binary coronary artery mask from an X-ray angiography frame using a transformer encoder, segmentation decoder, and MoE adapter parameters.The encoder and decoder use parameters θE and θD, while θA denotes the proposed MoE-based feature adapter parameters.
- Network formulation: The encoder extracts hierarchical features, MoE adapters refine selected intermediate token features, and the decoder fuses multi-scale representations for vessel probability prediction.This encoder-decoder design supports final binary vessel probability map generation.
- MoE feature adapter: Each MoE adapter replaces a single fixed branch with multiple lightweight experts and a router that activates only the top-k experts for an adaptive residual update.The router predicts expert importance from each input feature, and the selected expert outputs are combined.
- MoE feature adapter: The lightweight adapter uses bottleneck projection and residual refinement, with Wdown mapping C channels to d dimensions where d ≪ C, and Wup restoring the original dimension.A non-linear activation σ is used, while λ controls the adapter branch contribution.
- Multi-stage adaptation: MoE adapters are inserted into multiple transformer stages to address coronary arteries spanning thick proximal segments and thin distal branches before decoder-based binary mask prediction.The selected insertion stages are represented by S, and adapted multi-level features are passed to the decoder.
- Training objective: Training combines binary Dice loss and binary cross-entropy loss, balancing foreground overlap with pixel-wise vessel and background classification supervision.The coefficient α balances the contribution of the BCE term.
3 Experiments
The experiments formulate binary coronary artery segmentation on MOSXAV and assess cross-dataset robustness by directly evaluating MOSXAV-trained models on XACV. Performance is measured with overlap and vessel-detection metrics, using a lightweight MoE adapter configuration and representative segmentation baselines.
- Datasets: MOSXAV is used for training, validation-based model selection, and testing, with vessel annotations as foreground and all other pixels as background.All models are trained on the MOSXAV training set, selected using validation data, and evaluated on its test set.
- Datasets: Models trained on MOSXAV are directly evaluated on XACV using binary vessel ground truth to assess cross-dataset robustness.XACV provides manually annotated vessel segmentation ground truth for external evaluation.
- Implementation: The MoE adapter uses M = 4 experts, activates top-k = 2 experts per feature, has bottleneck ratio 1, spans all ViT stages, and uses α = 0.7.These settings define the adapter’s expert capacity, routing, placement, and loss weighting.
- Baselines: The proposed method is compared with U-Net, Attention U-Net, nnU-Net, nnWNet, AdapterSeg, MaskVSC, and a validated pseudo-label-based diffusion segmentation model.The comparison includes representative convolutional, adapter-based, and diffusion-based baselines for binary vessel segmentation.
- Evaluation Metrics: Binary segmentation is evaluated with Dice, Intersection over Union (IoU), precision, and recall, while F1 is omitted because it is equivalent to Dice in this setting.Dice and IoU measure mask overlap; precision and recall assess false-positive suppression and vessel recovery.
4 Results
The proposed MoE feature adapter achieves strong validation and test performance on MOSXAV and generalises better to unseen XACV imaging distributions. Ablations show that adapter-based adaptation, multiple experts, and top-k routing each contribute to performance and efficiency.
- MOSXAV validation: On MOSXAV validation, the proposed method achieves the best Dice, IoU, and recall, while providing a better precision-recall balance than conservative nnU-Net and nnWNet predictions.nnU-Net and nnWNet obtain higher precision but lower recall.
- MOSXAV test: On MOSXAV test, the proposed method achieves the best Dice and IoU and highest precision, outperforming AdapterSeg despite AdapterSeg’s higher recall.AdapterSeg’s lower precision indicates more false-positive predictions.
- External evaluation: On external XACV testing without fine-tuning, the proposed method achieves the best Dice, IoU, and recall, demonstrating stronger cross-dataset generalisation than the compared baselines.nnWNet has slightly higher precision, but lower recall than the proposed method.
- External evaluation: The MoE-based adapter improves robustness under unseen imaging distributions rather than merely overfitting MOSXAV, despite differences in vessel contrast, backgrounds, and acquisition characteristics.The external setting changes imaging conditions relative to training data.
- Ablation study: Ablations show that a single adapter substantially improves Dice and IoU over the backbone alone, while the MoE adapter further improves performance through complementary expert pathways.Top-k routing achieves the best Dice, IoU, precision, and recall, balancing expert specialisation and feature diversity better than top-1 or all-expert activation.
5 Conclusion
The paper presents a prompt-free MoE-based feature adapter for binary coronary artery segmentation in X-ray angiography videos, using multiple experts and input-dependent top-k routing. Experiments show improved segmentation and cross-dataset generalisation, while future work will investigate temporal modelling and downstream clinical measurements.
- 5 Conclusion: The proposed framework extends AdaptFormer-style lightweight adapters with multiple expert branches and input-dependent top-k routing for adaptive feature refinement.The design targets heterogeneous angiographic appearances while remaining prompt-free for binary coronary artery segmentation.
- 5 Conclusion: Experiments on MOSXAV and external testing on XACV demonstrate improved segmentation performance and stronger cross-dataset generalisation versus representative baselines.The reported evaluation supports the method’s robustness for coronary artery segmentation in X-ray angiography videos.
- 5 Conclusion: Future work will explore temporal modelling across angiography videos and its impact on vessel diameter estimation and stenosis quantification.These downstream clinical measurements are identified as targets for evaluating the value of temporal modelling.