Sparse Representation Learning for Vessels

Technical Summary: VAEsselSparse for Sparse Representation Learning of Vessels

Problem Statement
Analyzing human vasculature and vessel-like structures (e.g., airways) at sub-millimeter resolution is critical for clinical diagnosis, risk stratification, and treatment planning. However, extracting these networks from imaging modalities results in large 3D segmentation maps. Learning effective representations for such high-resolution volumetric data poses significant computational challenges. Existing methods often rely on simplifying assumptions—such as modeling vessels as tree-only structures, splines, or metric graphs—which sacrifice essential 3D geometric details or limit analysis to small regions. Furthermore, standard patch-based representation learning strategies fail to capture the global context of entire organ-level vessel networks. Dense convolutional approaches struggle with the computational cost of processing full organ-level volumes at high resolutions, while often losing fine structural connectivity due to averaging effects.

Methodology: VAEsselSparse
The authors propose VAEsselSparse, an efficient encoder-decoder architecture designed to learn compact, meaningful representations of large-scale, organ-level 3D vascular networks directly from sparse segmentation masks.

• Sparse Formulation: Unlike dense 3D images, vascular segmentation maps are inherently sparse ($\|x\|_{\ell_0} \ll H \cdot W \cdot D$). VAEsselSparse represents the input not as a dense grid but as a sparse set of active voxel coordinates $C_x$ and their associated feature vectors, forming a sparse tensor $X = (C_x, F_x)$. This allows computation to scale with the number of active voxels rather than the total volume size.
• Architecture: The model employs a Variational Autoencoder (VAE) framework utilizing sparse convolutions and sparse attention mechanisms.
  • Encoder ($E_\phi$): Local geometry is extracted using residual sparse 3D CNN blocks (sparsity-preserving convolutions, group normalization, ReLU). The spatial resolution is downsampled via strided sparse convolutions to reach the target latent resolution. To preserve global connectivity often lost in aggressive compression, the bottleneck voxels are treated as tokens with coordinate-based positional encoding and processed through sparse windowed self-attention layers. This enables information flow across distant vessel segments while maintaining tractable computational costs.
  • Decoder ($D_\theta$): Mirrors the encoder, refining latent tokens with sparse transformer blocks before progressively upsampling the representation using transposed sparse convolutions and residual blocks. The final output is a probabilistic reconstruction via a sigmoid activation.
• Training Objective: The model is trained with a standard VAE objective comprising a reconstruction term and a KL regularizer. Crucially, the reconstruction loss is computed only on the union of active voxels from the ground truth and the reconstruction support, avoiding dense supervision over the entire grid.
• Compression: The architecture achieves a spatial compression ratio of $8 \times 8 \times 8$ (and a volumetric compression ratio dependent on channel depth), significantly higher than previous volumetric VAEs, while maintaining structural connectivity.

Key Contributions

1. Architecture Design: Introduction of VAEsselSparse, an encoder-decoder tailored for sparse 3D vascular structures that combines sparse convolutions and sparse attention to handle organ-level networks at sub-millimeter resolution without simplifying assumptions.
2. Performance: Demonstration of superior reconstruction performance compared to dense counterparts and previous methods, achieving high spatial compression ($8 \times 8 \times 8$) while preserving topological connectivity.
3. Downstream Utility: Validation that the learned latent space retains clinically relevant discriminative features, effectively supporting downstream tasks such as vascular classification (e.g., aneurysm/stenosis detection, Circle of Willis subvariants) and serving as an embedding space for generative modeling to synthesize realistic vasculature.

Experimental Results
The method was evaluated on a consolidated dataset of airway, cerebral, and pulmonary vessels (isotropic 0.5 mm resolution) and specific challenge datasets (INSTED, TopCoW).

• Reconstruction: VAEsselSparse outperformed a dense VAE baseline (with similar parameter counts) and a state-of-the-art MedVAE (which used lower $4 \times 4 \times 4$ compression) across metrics including Dice, clDice, and Betti numbers (topological errors). Specifically, it achieved a clDice of 92.11 compared to 73.48 for the dense VAE at the same $8 \times 8 \times 8$ compression, and significantly outperformed VesselGPT (clDice 1.64) on complex structures.
• Classification: On unseen datasets (INSTED and TopCoW), the latent space enabled strong classification performance. An MLP trained on VAEsselSparse features achieved 87.96% balanced accuracy on INSTED (vs. 85.24% for ResNet and 34.89% for Dense VAE+PCA+RF). t-SNE visualizations confirmed that VAEsselSparse latent spaces form distinct clusters for healthy, aneurysm, and stenosis classes, whereas dense VAEs failed to separate these classes.
• Generative Modeling: A denoising U-Net trained on the VAEsselSparse latent space produced more realistic vessel samples than one trained on the dense VAE latent space. Quantitative metrics (Minimum Matching Distance, Coverage, 1-Nearest Neighbor Accuracy) favored VAEsselSparse, attributed to its sparse, low-dimensional latent space which mitigates issues related to high dimensionality and limited sample sizes.

Significance and Claims
The paper claims that VAEsselSparse addresses a critical gap in existing methods by enabling the learning of meaningful representations for entire organ-level vascular networks at clinical (sub-millimeter) resolution without relying on geometric simplifications. By leveraging the inherent sparsity of vessel structures, the method achieves substantial computational efficiency and spatial compression. The authors posit that this sparsity-preserving approach provides a practical path toward scalable, clinically useful modeling of high-resolution vasculature, facilitating both accurate reconstruction and the synthesis of realistic vascular structures for downstream clinical applications.