[Paper Review] Multi-View Deformable Convolution Meets Visual Mamba for Coronary Artery Segmentation

Xiaochan Yuan, Pai Zeng|arXiv (Cornell University)|Mar 23, 2026

Retinal Imaging and Analysis0 citations

TL;DR

The paper introduces MDSVM-UNet, a two-stage coronary artery segmentation model combining Multidirectional Snake Convolution (MDSConv) and Residual Visual Mamba (RVM), achieving state-of-the-art results on ImageCAS.

ABSTRACT

Accurate segmentation of coronary arteries from computed tomography angiography (CTA) images is of paramount clinical importance for the diagnosis and treatment planning of cardiovascular diseases. However, coronary artery segmentation remains challenging due to the inherent multi-branching and slender tubular morphology of the vasculature, compounded by severe class imbalance between foreground vessels and background tissue. Conventional convolutional neural network (CNN)-based approaches struggle to capture long-range dependencies among spatially distant vascular structures, while Vision Transformer (ViT)-based methods incur prohibitive computational overhead that hinders deployment in resource-constrained clinical settings. Motivated by the recent success of state space models (SSMs) in efficiently modeling long-range sequential dependencies with linear complexity, we propose MDSVM-UNet, a novel two-stage coronary artery segmentation framework that synergistically integrates multidirectional snake convolution (MDSConv) with residual visual Mamba (RVM). In the encoding stage, we introduce MDSConv, a deformable convolution module that learns adaptive offsets along three orthogonal anatomical planes -- sagittal, coronal, and axial -- thereby enabling comprehensive multi-view feature fusion that faithfully captures the elongated and tortuous geometry of coronary vessels. In the decoding stage, we design an RVM-based upsampling decoder block that leverages selective state space mechanisms to model inter-slice long-range dependencies while preserving linear computational complexity. Furthermore, we propose a progressive two-stage segmentation strategy: the first stage performs coarse whole-image segmentation to guide intelligent block extraction, while the second stage conducts fine-grained block-level segmentation to recover vascular details and suppress false positives..

Motivation & Objective

• Motivate automated, accurate coronary artery segmentation from CTA to aid diagnosis and treatment planning.
• Address the tubular, elongated vascular morphology and severe foreground-background imbalance.
• Propose a two-stage coarse-to-fine framework that balances global context and local vascular detail.

Proposed method

• Introduce MDSConv that performs deformable convolution along three orthogonal anatomical planes (sagittal, coronal, axial) and fuses multi-view features for tubular structures.
• Use a Residual Visual Mamba (RVM) decoder to model inter-slice long-range dependencies with linear complexity.
• Adopt a two-stage progressive segmentation: coarse whole-image segmentation to guide block extraction, followed by fine-grained block-level segmentation.
• Employ UNet++-style dense skip connections to enable rich multi-scale feature propagation.
• Train with Dice loss to handle severe class imbalance between vessels and background.

Experimental results

Research questions

• RQ1 Can multidirectional deformable convolution better capture the elongated and tortuous coronary vasculature than standard CNNs?
• RQ2 Does integrating a residual Visual Mamba decoder improve long-range dependency modeling and segmentation quality with linear complexity?
• RQ3 Does a two-stage coarse-to-fine strategy improve vascular continuity and reduce false positives compared to single-stage methods?
• RQ4 How does MDSVM-UNet perform on the ImageCAS benchmark relative to state-of-the-art methods in terms of DSC, HD, and AHD?

Key findings

• MDSVM-UNet achieves 0.686 DSC in single-stage and 0.8365 DSC in two-stage configuration on ImageCAS (with Stage 2 Dice loss).
• Two-stage MDSVM-UNet outperforms the ImageCAS baseline by 5.41% in DSC, 8.5456 in HD, and 0.8093 in AHD.
• Stage 1 (single-stage) DSC of 0.6860 with Dice loss and HD 27.8430; AHD 0.9023 (Table 1).
• Two-stage MDSVM-UNet with Dice loss achieves DSC 0.8365 and HD 27.8430, AHD 0.9023 (Table 2).
• The model has 26.7M parameters, offering competitive efficiency vs. transformer-based approaches.