Technical Summary: Deep Spectral Models for Robust Dental Shape Generation (ToothForge)

1. Problem Statement

Accurate modeling of dental crown morphology is essential for diagnosis, orthodontic planning, and computer-aided restoration design. However, developing robust generative models for this domain faces two primary challenges:

1. Data Scarcity and Dimensionality: Medical datasets, particularly in digital dentistry, are often small due to privacy constraints and costly annotations. Yet, individual samples (3D meshes) contain tens or hundreds of thousands of vertices. This creates a severe imbalance between data size and representational dimensionality, leading to overfitting risks in high-resolution spatial models.
2. Connectivity Inconsistency: Traditional deep learning approaches operating on spatial coordinates (e.g., point clouds or meshes) often require consistent mesh connectivity or point-wise correspondences across the dataset. In clinical practice, shapes from different clinics or scanners possess heterogeneous triangulations and vertex counts, making standard alignment difficult.
3. Limitations of Existing Methods:
   • Statistical Shape Models (SSMs/PCA): While compact and interpretable, they rely on linear assumptions that fail to capture complex, non-linear anatomical variations (e.g., non-convex occlusal surfaces).
   • Spatial Deep Generative Models (VAEs/Diffusion): These can model non-linearities but struggle with scalability on high-resolution data and small datasets. They often require expensive neighborhood searches and produce unordered point sets lacking explicit connectivity, necessitating post-processing that may distort anatomy.
   • Spectral Approaches: Previous spectral methods (e.g., Lemeunier et al., 2022) require identical mesh connectivity across all training samples, rendering them unsuitable for heterogeneous clinical datasets.

2. Methodology: ToothForge

The paper proposes ToothForge, a deep spectral generative framework designed to model dental crown geometries from compact, intrinsic representations without requiring consistent mesh connectivity.

Core Components

• Spectral Representation: Instead of operating on raw vertex coordinates, ToothForge projects 3D meshes into the eigenspace of the Laplace-Beltrami operator. A mesh is represented by a truncated set of spectral coefficients ($C_k = \Phi_k^\top V$), where low frequencies capture global form and high frequencies encode fine details. This provides an intrinsic, ordered, and compact description of the geometry.
• Spectral Synchronization: To address the instability of Laplace-Beltrami eigenbases (where eigenfunctions may flip signs or swap order across shapes), the framework employs spectral synchronization. All shapes are aligned to a common reference basis ($\Phi_{k,ref}$) via a transformation matrix ($R_i$). This ensures consistent spectral embeddings across shapes with arbitrary connectivity, effectively removing decomposition bias without deforming the shapes toward a geometric prior.
• Generative Architecture: The aligned spectral coefficients are modeled using a $\beta$-Variational Autoencoder ($\beta$-VAE).
  • Encoder/Decoder: A 5-stage 1D convolutional architecture processes the ordered spectral coefficients.
  • Training Objective: The loss function combines the Mean Squared Error (MSE) of the spectral coefficients with a KL-divergence term to regularize the latent space. Crucially, optimization occurs entirely in the spectral domain; no spatial regularization is required.
• Generation Process: New shapes are synthesized by sampling latent vectors from a Gaussian prior, decoding them into spectral coefficients, and projecting them back to the spatial domain using the common reference eigenbasis and the template mesh connectivity. This guarantees point-to-point correspondence across all generated samples.

Baselines

The framework is benchmarked against:

1. PCA-based SSM: A linear baseline requiring dense correspondence.
2. PointVAE: A spatial deep generative model operating on unordered point clouds using PointNet++ architectures.
3. PointDiffusion: A point-based diffusion probabilistic model.

3. Key Contributions

The authors present the following contributions:

• A Connectivity-Independent Framework: ToothForge enables deep generative modeling on datasets with variable triangulations by leveraging synchronized spectral embeddings, overcoming the connectivity constraints of prior spectral methods.
• Spectral Sufficiency: The study provides evidence that latent regularization performed entirely in the spectral domain is sufficient for effective learning. Adding spatial regularization terms (e.g., Chamfer distance on vertices) provided no significant benefit, reinforcing that intrinsic spectral representations are adequate for modeling anatomical variability.
• Comprehensive Robustness Analysis: The framework is evaluated regarding its robustness to the choice of reference template, the number of spectral modes ($k$), and different tooth classes (incisors, premolars, molars).
• Rigorous Benchmarking: Direct comparisons are made against PCA, PointVAE, and PointDiffusion, supported by ablation studies and downstream task evaluations.

4. Results

Experiments were conducted on a private dataset of 430 dental crowns (149 incisors, 161 premolars, 120 molars).

• Reconstruction Quality: ToothForge achieves reconstruction errors comparable to or exceeding spatial approaches. Absolute reconstruction errors remain low across all tooth classes and spectral truncation levels ($k$). While relative errors increase slightly with higher $k$ (due to noise in high-frequency components), the model maintains high fidelity to the ground truth.
• Generative Performance:
  • MMD (Minimum Matching Distance): ToothForge achieves the lowest MMD scores (e.g., 0.0997 for molars), indicating generated samples are highly faithful to the training distribution.
  • Coverage: The model demonstrates strong coverage (43.78% for molars), comparable to PointVAE and PointDiffusion.
  • Qualitative Analysis: Unlike PCA, which produces overly smooth shapes lacking fine anatomical details (e.g., lingual/buccal grooves), ToothForge captures a broader range of morphological variations. Unlike point-based methods, it produces coherent, watertight surfaces directly without requiring post-processing surface reconstruction, which often introduces noise or distortion.
• Efficiency: ToothForge is significantly faster in sampling (milliseconds) compared to PointDiffusion (seconds) and requires less computational time for training than diffusion models.
• Ablation Studies:
  • Removing spectral synchronization ("No spec. sync.") resulted in implausible shapes and significantly higher errors.
  • Removing latent regularization ($\beta=0$) caused the model to collapse to mean-like geometries with no variability.
  • Adding spatial regularization terms did not improve performance, confirming the sufficiency of spectral learning.
• Downstream Task (Data Augmentation): In a low-data tooth classification task (100 training samples), ToothForge-generated samples significantly improved classifier accuracy (98.79% with 1000 samples vs. 90.36% with no augmentation). Combining ToothForge augmentation with standard mesh transforms yielded the best results, suggesting the generated shapes provide non-trivial, complementary variability.

5. Significance and Claims

The paper claims that ToothForge offers a practical solution for medical shape generation where datasets are small and consistent connectivity cannot be guaranteed.

• Clinical Applicability: By decoupling the learning space dimensionality from mesh resolution and handling heterogeneous connectivities, the framework is particularly suitable for real-world clinical scenarios where shapes from various clinics cannot be guaranteed to share the same topology.
• Balance of Properties: ToothForge successfully balances the compactness and interpretability of spectral methods with the generative power of non-linear deep learning. It avoids the computational burden and instability of high-dimensional spatial models while capturing non-linear anatomical variations that linear SSMs miss.
• Future Directions: The authors suggest that this framework paves the way for unified latent spaces modeling multiple tooth classes jointly, conditional generation (e.g., guided by target crown size), and controllable shape editing workflows. They explicitly state that the primary value lies not just in improved reconstruction over PCA under ideal conditions, but in providing a robust generative representation for clinical datasets with inconsistent discretization.

The authors conclude that synchronized spectral modeling is a robust, efficient, and geometrically interpretable approach for biomedical imaging, specifically addressing the constraints of limited data and high-resolution representation in digital dentistry.