Introduction
Accurate 3D segmentation of vascular structures—such as those in the brain, heart, or aorta—is a cornerstone of modern medical imaging and disease diagnosis. However, the development of robust deep learning models for this task is severely bottlenecked by the requirement for "full supervision." In a fully supervised setting, radiologists must manually label every vessel voxel within a 3D volume. Given that vessels are thin, sparse, and follow complex, tree-like paths, this process is notoriously labor-intensive, error-prone, and slow.
While weak supervision methods (using points, scribbles, or bounding boxes) have been proposed to alleviate this burden, they often struggle to capture the intricate connectivity of vascular networks. This paper introduces a novel weakly-supervised framework that utilizes Maximum Intensity Projection (MIP) as a proxy for 3D annotation. By shifting the labeling task from 3D volumes to 2D MIP images, the authors achieve performance comparable to full supervision while reducing the annotation workload by over 75%.
The MIP-Based Annotation Strategy
Maximum Intensity Projection (MIP) is a visualization technique that flattens a 3D volume into a 2D image by selecting the voxel with the highest intensity along each projection ray. For vascular imaging, MIP is particularly effective because it preserves the continuity of the vessel tree, making it much more intuitive for human annotators than individual 2D slices.
The proposed pipeline begins with radiologists providing 2D annotations on these MIP images. To translate these 2D labels back into the 3D domain, the system utilizes an index map, which records the specific 3D coordinates of the voxel that contributed to each 2D pixel. These back-projected points serve as seeds for a region-growing algorithm, generating initial 3D pseudo-labels that serve as the starting point for model training.
2D-3D Feature Fusion and Refinement
To bridge the gap between 2D supervision and 3D segmentation, the authors designed a specialized 2D-3D Feature Fusion Network. The architecture consists of two primary branches:
- 3D Branch: A standard 3D U-Net that processes the volumetric data.
- 2D Branch: A 2D U-Net that processes the MIP image and its corresponding 2D ground truth labels.
Crucially, the network uses the index map to project features from the 3D layers into the 2D layers. This allows the 2D branch—which has access to high-quality, human-verified labels—to provide "oversight" and guidance to the 3D branch. This fusion ensures that the 3D model learns to respect the connectivity and boundaries established in the 2D projection.
To address the inherent noise in initial pseudo-labels, the framework incorporates two refinement strategies:
- Confident Learning (CL): This identifies and removes potentially noisy or incorrect labels from the training set.
- Uncertainty Estimation (UE): Utilizing Monte Carlo dropout, the system identifies reliable unlabeled voxels and incorporates them into the training process, iteratively improving the model's robustness.
Experimental Results
The framework was validated across five diverse datasets, including MRA and CTA images of the brain, heart, and aorta. The results demonstrate that the method is highly competitive with fully supervised approaches:
- Accuracy: On the Cerebral MRA dataset, the proposed method achieved a Dice score (DSC) of 84.35%, nearly matching the fully supervised upper bound of 85.07%.
- Efficiency: The method required only 7.9 hours of total annotation time to reach high accuracy, compared to 36.7 hours for full 3D supervision—a nearly 5x improvement in efficiency.
- Connectivity: Using the ClDice metric, the method significantly outperformed other weak-labeling strategies (like scribbles or points), proving its superior ability to maintain the structural integrity of the vascular tree.
Discussion and Limitations
The primary insight of this work is that MIP images serve as an ideal proxy for 3D vascular structures. By using the index map as a mathematical bridge, the authors successfully transferred the ease of 2D labeling to the complexity of 3D segmentation. This suggests that for a fixed "time budget," it is more effective to collect a large volume of 2D MIP-labeled data than a small amount of perfectly labeled 3D data.
However, challenges remain. In certain CTA views, high-intensity structures like the skull or metal implants can occlude vessels in the MIP projection. While preprocessing steps like skull stripping help, future work will need to address these occlusions more robustly. Additionally, while 2D labeling is significantly faster, it still requires manual effort; the authors suggest that future iterations could incorporate self-supervised learning to further automate the 2D annotation phase.
Conclusion
This research presents a practical and highly efficient solution to the 3D segmentation bottleneck. By leveraging the intuitive nature of MIP images and a sophisticated 2D-3D fusion architecture, the framework enables the development of high-performance vascular segmentation models with a fraction of the manual labor previously required.
Summary updated