Abstract
Medical image analysis, particularly 3D carotid artery segmentation, is often hindered by sparse annotation data. Existing datasets frequently contain spatially discontinuous slice annotations, where only a small fraction of the volumetric data is expert-labeled. To overcome this, we introduce DBF-UNet, a two-stage framework that leverages foundation models for pseudo-label generation and a novel lightweight 3D architecture for precise vascular segmentation. Experimental results on the COSMOS2022 dataset show that our method achieves superior performance (95.22% Dice for lumen) with significantly fewer parameters (2.81M) than existing state-of-the-art models.
Introduction
In 3D MRI scans of the carotid artery, manual annotation is labor-intensive and requires specialized expertise. Consequently, many research datasets are "spatially discontinuous," meaning experts only label a subset of slices within a 3D volume. This sparsity creates a significant training gap for standard 3D segmentation models like nnUNet, which struggle to learn continuous vascular structures from missing data. Furthermore, simple interpolation between labeled slices often results in "staircase" artifacts and fails to capture complex morphologies at vessel bifurcations.
Methodology
Stage 1: SAM-Refined Pseudo-Labeling (S-RPL)
The first stage of our framework addresses the annotation gap by generating high-quality pseudo-labels for unlabeled slices. We first construct continuous vessel centerlines by interpolating between the centroids of expert-annotated slices. Labels are then propagated along these centerlines to create initial "weak" annotations (Centroid-guided Interpolated Pseudo Labels, or C-IPL).
To refine these labels into "strong" supervision, we utilize SAM-Med2D, a medical adaptation of the Segment Anything Model. We fine-tune SAM-Med2D on the available expert slices. During inference on unlabeled slices, we employ a noise-based prompting strategy: multiple bounding boxes are generated with random shifts and fed into the model. An ensemble voting mechanism then aggregates the resulting masks to produce the final SAM-Refined Pseudo-Label (S-RPL). This process ensures that the generated labels are anatomically consistent even when the initial interpolation is imperfect.
Stage 2: DBF-UNet Architecture
With the dataset completed via S-RPL, we train the Dense Bidirectional Feature Fusion UNet (DBF-UNet). This architecture is designed for high efficiency and effective feature reuse, comprising three main components:
- DSDBlock: Utilizes dense connectivity and depthwise separable convolutions for efficient downsampling.
- MLKBlock: A multi-level kernel block that uses varying kernel sizes (3x3, 5x5, 7x7) and statistical attention to capture features across multiple scales.
- BFFBlock: A bidirectional feature fusion block that enables information flow both upwards and downwards between encoder levels, preserving fine-grained spatial details crucial for thin vessel walls.
Results and Discussion
We validated our framework on the COSMOS2022 dataset, consisting of 3D black-blood vessel wall MRI scans. The impact of the S-RPL stage is clearly visible in the 3D reconstructions of the carotid artery. While standard interpolation methods produce jagged, discontinuous structures, the SAM-refined labels result in a smooth, continuous vascular geometry that closely mirrors real anatomy.
Quantitatively, DBF-UNet outperforms several heavyweights in the field:
- Segmentation Accuracy: It achieved a 95.22% Dice score for vessel lumen and 86.08% for vessel walls, surpassing nnUNet (92.52%) and Transformer-based models like SwinUNETR (90.86%).
- Computational Efficiency: DBF-UNet uses only 2.81 million parameters. In contrast, SwinUNETR uses 15.64M, and UNETR uses 121.35M. This represents a nearly 40x reduction in parameter count compared to the largest models while delivering superior accuracy.
Conclusion
DBF-UNet provides a robust solution to the sparse annotation problem in 3D medical imaging. By using foundation models as intelligent "labelers" in a two-stage process, we can transform weak, interpolated data into high-quality supervision. The lightweight DBF-UNet architecture further proves that targeted feature fusion and multi-scale aggregation can outperform massive models, making high-precision 3D carotid artery segmentation more accessible and computationally efficient.
The source code for this project is available at https://github.com/Haoxuanli-Thu/DBF-UNet.
Summary updated