Abstract
Automated segmentation of cerebral vasculature in Digital Subtraction Angiography (DSA) is critical for stroke and aneurysm treatment but is hindered by the high cost of expert manual annotation. This paper presents a weakly-supervised framework that utilizes Active Contour Models (ACM) to generate training labels. By incorporating a human-in-the-loop selection process (HACWA), the method achieves a 100x reduction in annotation time. Experimental results demonstrate that models trained on these weak labels, supplemented by a small fraction of manual annotations, outperform human experts, achieving a DICE score of 84.0%.
Introduction
Digital Subtraction Angiography (DSA) remains the gold standard for visualizing intracranial blood vessels. However, the intricate nature of these structures makes manual segmentation extremely labor-intensive, often requiring nearly two hours per image. Standard deep learning approaches typically require large, high-quality annotated datasets, which are prohibitively expensive to produce in the medical domain. This research addresses the data labeling bottleneck by proposing a weakly-supervised pipeline that leverages automated "silver standard" labels to train high-performing convolutional neural networks (CNNs).
Methods
The proposed framework employs Active Contour Models (ACM) to generate initial vessel masks. Two primary strategies were evaluated:
- Fully-automated Active Contour Weak Annotations (FACWA): Labels are generated autonomously using fixed parameters.
- Human-in-the-loop Active Contour Weak Annotations (HACWA): The system generates multiple mask candidates using varying parameters, and a human expert selects the best version. This selection process takes approximately one minute, compared to nearly two hours for manual drawing.
The training pipeline uses a DenseNet169 encoder paired with a Feature Pyramid Network (FPN) decoder. The researchers also explored hybrid training, combining weak labels with varying percentages of manual annotations (MA) to optimize the performance-to-effort ratio.
Results
The framework was evaluated on a new dataset of 128 high-quality cerebral DSA images and the benchmark DRIVE retinal dataset. The results indicate that "good enough" labels can produce superior models. Specifically, the 100% HACWA + 11% MA configuration outperformed a second human annotator (82.5% vs 81.1% DICE). The peak performance (84.0% DICE) was achieved using 100% HACWA and 48% MA, significantly reducing the total human effort required.
Discussion and Conclusion
The study demonstrates that human expertise is more efficiently utilized in a selection role rather than a generative one for medical image annotation. While active contours are sensitive to noise, the downstream CNNs effectively learn to generalize beyond the imperfections in the weak labels. Future work will focus on the robustness of this approach across diverse imaging hardware and pathologies. The release of the cerebral DSA dataset provides a valuable benchmark for the medical imaging community.
Summary updated