Bone Occupancy Function in 3D Modeling

• Bone occupancy function is a continuous representation that assigns a scalar value in [0,1] to indicate the probability of a point being within bone.
• It leverages neural implicit methods and coordinate-based MLPs to enable high-fidelity, resolution-independent 3D reconstruction from sparse clinical imaging.
• Hybrid approaches integrate template-guided mesh refinement and classical level-set methods to ensure anatomical plausibility and simulate dynamic bone remodeling.

The bone occupancy function encodes the spatial presence of bone in a given anatomical region, assigning to each point in $\mathbb{R}^3$ a scalar value—typically in $[0,1]$—that expresses either the probability (in learned models) or the binary state (in analytic models) of the point being inside bone. This continuous representation underpins diverse methodologies for 3D bone reconstruction from sparse clinical imaging, morphological analysis, and computational modeling of biological processes such as bone remodeling. Recent research utilizes neural implicit methods, coordinate-based multi-layer perceptrons (MLPs), and level-set techniques to construct high-fidelity, resolution-independent bone occupancy fields for both data-driven reconstruction and mechanobiological simulations (Guven et al., 11 Apr 2025, Chen et al., 2024, Graham et al., 2011, Mihajlovic et al., 2021).

1. Mathematical Formulation and Interpretation

The bone occupancy function $f: \mathbb{R}^3\to [0,1]$ is defined such that, for an arbitrary spatial point $x$, $f(x)\approx 1$ if $x$ lies inside bone and $f(x)\approx 0$ otherwise. In learned implicit settings, $f$ may be parameterized by neural network weights, integrating both spatial coordinates and, optionally, conditioning image features or pose parameters. In classical frameworks, such as morphogenesis or remodeling models, a level-set function $\varphi(x, t)$ partitions space sharply:

• $\varphi(x, t) > 0$: $[0,1]$0 is bone,
• $[0,1]$1: $[0,1]$2 is non-bone (e.g., marrow),
• $[0,1]$3: $[0,1]$4 is on the interface (the bone surface) (Graham et al., 2011).

The bone surface is the set of points $[0,1]$5, where $[0,1]$6 is a threshold parameter, commonly $[0,1]$7 or, in some models, $[0,1]$8 (Guven et al., 11 Apr 2025, Chen et al., 2024).

2. Neural Implicit Occupancy Modeling

The neural implicit paradigm models $[0,1]$9 using a coordinate-conditioned MLP, optionally conditioned on global or local image-derived context codes. For example, in X2BR, the network $f: \mathbb{R}^3\to [0,1]$0 (where $f: \mathbb{R}^3\to [0,1]$1 is a DRR image) is realized as follows (Guven et al., 11 Apr 2025):

• Encoder: ConvNeXt infers a global code $f: \mathbb{R}^3\to [0,1]$2 from input X-ray data.
• Decoder: An MLP $f: \mathbb{R}^3\to [0,1]$3 receives the concatenation $f: \mathbb{R}^3\to [0,1]$4 and outputs $f: \mathbb{R}^3\to [0,1]$5, using DenseNet-inspired architectures interleaved with Conditional Batch Normalization (CBN), which modulates hidden activations by the conditioning code.

Other frameworks, e.g., SdAOF, employ multi-view fusion and a minimal MLP, drawing feature vectors from per-view CNNs at each query point via projective mapping, then estimating occupancy as $f: \mathbb{R}^3\to [0,1]$6 for arbitrary $f: \mathbb{R}^3\to [0,1]$7 (Chen et al., 2024). Articulated human models such as LEAP generalize $f: \mathbb{R}^3\to [0,1]$8 to $f: \mathbb{R}^3\to [0,1]$9, conditioning via hierarchical skinning weights, canonicalization, and composite local and global encodings of articulated bone structure (Mihajlovic et al., 2021).

3. Training Objectives and Supervision

Neural bone occupancy functions are commonly supervised using pointwise binary labels $x$0, typically drawn from ground-truth CT meshes:

• Losses: Standard choices include binary cross-entropy (Guven et al., 11 Apr 2025) and mean squared error (Chen et al., 2024, Mihajlovic et al., 2021), applied across batches of sampled points.
• Sampling: Points are drawn uniformly from the object's bounding volume or with emphasis near the surface to improve resolution.
• Regularization: Some frameworks eschew additional regularizers (e.g., Eikonal, total variation), leveraging the strength of their architectures and hybrid refinements (Guven et al., 11 Apr 2025). In others, auxiliary features or cycle consistency penalties are used to ensure coherence (e.g., cycle-distance in LEAP) (Mihajlovic et al., 2021).

Teacher-student strategies with spatial-division distillation are also employed to guide learning about occlusion and local spatial context, as in SdAOF (Chen et al., 2024).

4. Hybrid Template Integration and Post-processing

To enforce anatomical plausibility beyond thresholded occupancy fields, hybrid methods incorporate template-guided mesh refinement (Guven et al., 11 Apr 2025):

• Template construction: Anatomical templates are assembled using biomechanical skeleton models (e.g., SKEL BSM), with patient-specific alignment via bone detection (e.g., YOLOv9).
• Non-rigid registration: Occupancy-field surfaces are aligned to these templates using algorithms such as geodesic-based coherent point drift (GBCPD++), which optimize Gaussian-kernel-based deformation fields regularized by geodesic adjacency.
• Mesh fusion: The final output is produced by merging the deformed template mesh and the mesh extracted from the learned occupancy isosurface, often via Poisson blending or union. This step improves the anatomical consistency of reconstructed shape, notably for complex elements such as ribs and vertebrae.

A trade-off emerges between numerical fidelity (IoU, Chamfer distance) and anatomical realism, with hybrid models offering robust anatomical structure at only mild expense to raw surrogate metrics (Guven et al., 11 Apr 2025).

5. Classical Level-Set Formulation in Bone Biology

In bone biology, the bone occupancy function is formalized via level-set methods to model dynamic remodeling (Graham et al., 2011):

• Occupancy function: $x$1 partitions bone/marrow domains, with $x$2 denoting bone.
• Interface evolution: The surface $x$3 evolves by

$x$4

where $x$5 depends on local cellular activity (osteoclast and osteoblast dynamics via power-law ODEs), yielding normal velocity $x$6 for surface displacement.

• Boundary/initial conditions: Usually periodic or spatially homogeneous; cellular recruitment is rapid and restricted to the interface.

This approach enables mechanistic simulation of spatiotemporal bone remodeling, capturing complex interface movements without explicit surface tracking, and is extensible to include more comprehensive biological signaling (Graham et al., 2011).

6. Applications and Quantitative Performance

Bone occupancy functions are fundamental to state-of-the-art clinical and preclinical 3D bone reconstruction from sparse radiographs. Notable applications and evaluation outcomes:

• Single-view X-ray reconstruction (X2BR): Implicit occupancy yields high fidelity (IoU 0.952, Chamfer-L1 0.005) (Guven et al., 11 Apr 2025).
• Biplanar X-ray (SdAOF): Provides memory efficiency, arbitrary inference resolution, and occlusion-aware fine-scale surface recovery; achieves mean Chamfer Distance as low as 2.335 mm at high outputs (Chen et al., 2024).
• Articulated human models (LEAP): Enables generalization to unseen body shapes and poses (IoU up to 98.4% on DFaust), with differentiable occupancy facilitating physically plausible scene assembly (Mihajlovic et al., 2021).
• Biological simulation: Level-set bone occupancy functions support in silico experimentation on remodeling phenomena (Graham et al., 2011).

7. Methodological Variants and Extensions

Different bone occupancy approaches span purely coordinate-conditioned MLPs, multi-view feature fusion, skinning-based articulated representations, and physics-informed PDEs. Key architectural components include:

• Encoder backbones: ConvNeXt for global codes (Guven et al., 11 Apr 2025), 2D CNNs for pixel-aligned features (Chen et al., 2024).
• Decoders: MLPs modulated by conditioning vectors, with DenseNet or skip connections, often using CBN.
• Distillation and attention: Spatial-division augmented distillation to handle occlusion (Chen et al., 2024).
• Regularization: Use of anatomical priors and normalization to ensure scale-invariant learning and anatomically consistent output.

A plausible implication is that as data and computational resources increase, implicit and hybrid bone occupancy representations are likely to provide the backbone for real-time, error-bounded, and anatomically faithful 3D reconstructions in both clinical and research settings.

---

References:

• (Guven et al., 11 Apr 2025)
• (Chen et al., 2024)
• (Graham et al., 2011)
• (Mihajlovic et al., 2021)