DBF-UNet: A Two-Stage Framework for Carotid Artery Segmentation with Pseudo-Label Generation

TL;DR

This work tackles the challenge of sparse, spatially discontinuous annotations in 3D carotid artery segmentation by introducing a two-stage framework. It first generates high-quality pseudo-labels through centroid-guided interpolation and SAM-Med2D fine-tuning, then refines full 3D vessel segmentation with the proposed DBF-UNet, which features dense downsampling (DSDBlock), multi-scale feature learning (MLKBlock), and bidirectional feature fusion (BFFBlock). The approach achieves state-of-the-art results on the COSMOS2022 dataset, with lumen Dice of 95.22% and wall Dice of 86.08%, demonstrating strong performance in challenging regions and with limited annotations. This work highlights effective integration of foundation-model-based prompting with specialized lightweight architectures for precision, label-efficient medical image segmentation and has practical potential for clinical workflows.

Abstract

Medical image analysis faces significant challenges due to limited annotation data, particularly in three-dimensional carotid artery segmentation tasks, where existing datasets exhibit spatially discontinuous slice annotations with only a small portion of expert-labeled slices in complete 3D volumetric data. To address this challenge, we propose a two-stage segmentation framework. First, we construct continuous vessel centerlines by interpolating between annotated slice centroids and propagate labels along these centerlines to generate interpolated annotations for unlabeled slices. The slices with expert annotations are used for fine-tuning SAM-Med2D, while the interpolated labels on unlabeled slices serve as prompts to guide segmentation during inference. In the second stage, we propose a novel Dense Bidirectional Feature Fusion UNet (DBF-UNet). This lightweight architecture achieves precise segmentation of complete 3D vascular structures. The network incorporates bidirectional feature fusion in the encoder and integrates multi-scale feature aggregation with dense connectivity for effective feature reuse. Experimental validation on public datasets demonstrates that our proposed method effectively addresses the sparse annotation challenge in carotid artery segmentation while achieving superior performance compared to existing approaches. The source code is available at https://github.com/Haoxuanli-Thu/DBF-UNet.

Figures (8)

• Figure 1: Comparison of original label and three pseudo-label generation methods (A-IPL, C-IPL, and S-RPL).
• Figure 2: Original label and three pseudo-label generation approaches
• Figure 3: Simplified vessel wall segmentation workflow. Vessel lumen and wall annotations are merged into complete vessel masks (top). The vessel wall mask is obtained by subtracting the lumen mask from the complete vessel mask (bottom).
• Figure 4: Fine-tuning framework of SAM-Med2D. The architecture uses three prompt types: Points, Bboxes and Masks. Initially, a random foreground point or bounding box from C-IPL serves as the sparse prompt. In later iterations, the framework identifies error regions between Prediction and C-IPL to generate new prompts. Previous predictions as mask prompts in subsequent iterations. The process refines segmentation results over N-1 iterations.
• Figure 5: Inference optimization of SAM-Med2D