Skelite: Compact Neural Networks for Efficient Iterative Skeletonization

TL;DR

Skelite addresses the speed–topology trade-off in skeletonization by learning an iterative, differentiable skeletonization process trained on synthetic data. It uses a Bezier-based synthetic dataset, a neighborhood-based thinning loss, and a distillation framework to produce thin, connected skeletons with strong generalization to unseen domains. The approach delivers about a 100x speedup over topology-constrained baselines while maintaining competitive topological accuracy and improving connectivity in downstream segmentation via clDice. This yields an efficient, domain-robust skeleton prior applicable to 2D and 3D medical image analysis and downstream tasks.

Abstract

Skeletonization extracts thin representations from images that compactly encode their geometry and topology. These representations have become an important topological prior for preserving connectivity in curvilinear structures, aiding medical tasks like vessel segmentation. Existing compatible skeletonization algorithms face significant trade-offs: morphology-based approaches are computationally efficient but prone to frequent breakages, while topology-preserving methods require substantial computational resources. We propose a novel framework for training iterative skeletonization algorithms with a learnable component. The framework leverages synthetic data, task-specific augmentation, and a model distillation strategy to learn compact neural networks that produce thin, connected skeletons with a fully differentiable iterative algorithm. Our method demonstrates a 100 times speedup over topology-constrained algorithms while maintaining high accuracy and generalizing effectively to new domains without fine-tuning. Benchmarking and downstream validation in 2D and 3D tasks demonstrate its computational efficiency and real-world applicability

Figures (6)

• Figure 1: Existing skeletonization methods trade off speed, accuracy, and adaptability. Morphology-based methods often break skeletons, Boolean predicates are slow but topologically accurate, and U-Nets struggle with domain shifts. Our method is fast, adaptable, and produces thin skeletons with few breakages.
• Figure 2: Our neural network skeletonization module. The arrows indicate the number of output channels after each layer.
• Figure 3: Example of knowledge distillation strategy. The student network can receive very tight supervision at each skeletonization step from the teacher network.
• Figure 4: Top: Illustration of Bézier dataset generation. Bottom: representative samples in 2D and 3D.
• Figure 5: A Comparison of neural network skeletonization trained on the Bézier dataset and tested on DRIVE. U-Net-based methods overfit to Bézier-specific features, resulting in over-thinning in DRIVE. Skelite's controlled iterative thinning addresses this issue, allowing it to generalize to new domains more easily.