Abstract
Representing complex anatomical structures such as blood vessels and airways is a fundamental challenge in medical imaging. Traditional explicit representations like voxel grids are memory-intensive, while meshes often struggle with intricate branching topologies. We introduce TrIND (Trees by Implicit Neural Diffusion), a novel framework that leverages Implicit Neural Representations (INRs) to model anatomical trees. TrIND "overfits" a small Multi-Layer Perceptron (MLP) to each tree instance, acting as a continuous neural occupancy field. Generative modeling is then performed by training a denoising diffusion transformer directly on the flattened weight space of these MLPs. Our approach achieves up to 220× compression compared to volumetric grids and enables high-fidelity, resolution-independent synthesis of diverse anatomical structures. We demonstrate the versatility of TrIND across eight datasets, including intracranial vasculature and bronchial trees.
Introduction
Accurately representing anatomical trees, such as vasculature and bronchial networks, is critical for medical imaging, surgical planning, and fluid dynamics simulations. However, existing methods face significant trade-offs. Voxel grids suffer from $O(n^3)$ memory scaling, making high-resolution modeling prohibitive. Explicit meshes and medial axes often fail to capture the varying, intricate topologies of organic branching. Rule-based systems, like L-systems, lack the organic variability seen in real patient data.
TrIND addresses these issues by moving away from explicit geometry toward a continuous, coordinate-based representation. By modeling trees as neural fields and applying diffusion models to the parameters of these fields, we create a generative pipeline that is both memory-efficient and topologically robust.
Methodology
The TrIND framework consists of a two-stage pipeline: per-sample INR optimization followed by weight-space diffusion.
Stage 1: Learning Per-Sample INRs
Each anatomical tree is represented as an Implicit Neural Representation (INR). We "overfit" a small Multi-Layer Perceptron (MLP) to the occupancy field of an individual tree. Given a 3D coordinate $(x, y, z)$, the MLP outputs the probability of that point being inside the structure. This continuous formulation allows for arbitrary resolution querying, enabling the extraction of smooth surfaces via Marching Cubes without being bound to a fixed grid.
Stage 2: Diffusion in Weight Space
To generate novel structures, we perform generative modeling in the "weight space" of the trained MLPs. The weights and biases of the optimized networks are flattened into 1D vectors. We then train a Denoising Diffusion Model (DDM)—specifically a Diffusion Transformer—on these vectors. To accommodate the multi-layer structure of the MLPs, we decompose the weights into layer-specific tokens before processing them with the transformer. By reversing the diffusion process, the model samples new sets of MLP parameters that represent plausible, novel anatomical trees.
Results and Key Insights
The framework was evaluated across eight diverse datasets, including synthetic (VascuSynth), intracranial (IntRA, CoW), and clinical data (BraTS, DRIVE, EXACT, WBMRA).
- Massive Compression: INRs provide a powerful compression mechanism. TrIND achieved up to 220× compression relative to standard volumetric grids while maintaining higher reconstruction accuracy than decimated meshes of equivalent size.
- Biological Meaning in Weight Space: t-SNE visualizations reveal that the weight space of the MLPs preserves biological characteristics. Models representing trees with similar topological complexity (e.g., number of bifurcations) naturally cluster together.
- High-Fidelity Synthesis: TrIND consistently outperformed traditional downsampling methods in terms of Chamfer Distance. Generative quality, measured by Minimum Matching Distance (MMD) and Coverage (COV), confirmed that synthesized trees are both diverse and biologically plausible.
- Resolution Independence: The continuous nature of the representation allows the same model to generate low-resolution previews or high-fidelity meshes for specialized applications like fluid dynamics without retraining.
Conclusion
TrIND demonstrates that denoising diffusion on the parameters of implicit neural fields is a viable and efficient path for modeling complex 3D anatomical structures. While the current implementation requires significant training time (approx. 3 days on an A40 GPU), it offers a proof-of-concept for high-fidelity, compressed, and topologically aware generative modeling. Future work will focus on integrating semantic labeling into the neural fields and embedding TrIND into end-to-end medical segmentation and analysis pipelines.
Summary updated