NICE: Neural Implicit Craniofacial Model for Orthognathic Surgery Prediction

TL;DR

NICE tackles the challenge of predicting postoperative facial appearance after orthognathic surgery by combining neural implicit geometry with biomechanically informed deformation modeling. It introduces two modules: a shape module with region-specific SDF decoders to reconstruct face, maxilla, and mandible, and a surgery module with region-specific deformation decoders driven by a shared surgery latent code and anatomical priors to capture nonlinear skull–soft-tissue responses. The approach achieves superior geometric fidelity and regional accuracy (notably lips and chin) compared with state-of-the-art methods and remains clinically viable in terms of inference time and consistency. This work provides a principled, anatomically aware framework for planning and patient consultation in orthognathic procedures.

Abstract

Orthognathic surgery is a crucial intervention for correcting dentofacial skeletal deformities to enhance occlusal functionality and facial aesthetics. Accurate postoperative facial appearance prediction remains challenging due to the complex nonlinear interactions between skeletal movements and facial soft tissue. Existing biomechanical, parametric models and deep-learning approaches either lack computational efficiency or fail to fully capture these intricate interactions. To address these limitations, we propose Neural Implicit Craniofacial Model (NICE) which employs implicit neural representations for accurate anatomical reconstruction and surgical outcome prediction. NICE comprises a shape module, which employs region-specific implicit Signed Distance Function (SDF) decoders to reconstruct the facial surface, maxilla, and mandible, and a surgery module, which employs region-specific deformation decoders. These deformation decoders are driven by a shared surgical latent code to effectively model the complex, nonlinear biomechanical response of the facial surface to skeletal movements, incorporating anatomical prior knowledge. The deformation decoders output point-wise displacement fields, enabling precise modeling of surgical outcomes. Extensive experiments demonstrate that NICE outperforms current state-of-the-art methods, notably improving prediction accuracy in critical facial regions such as lips and chin, while robustly preserving anatomical integrity. This work provides a clinically viable tool for enhanced surgical planning and patient consultation in orthognathic procedures.

Figures (6)

• Figure 1: Pipeline of our neural implicit craniofacial model (NICE) for postoperative facial appearance prediction. The shape module first reconstructs the preoperative facial surface $\mathbf{F}_\text{face-pre}$ and skull $\mathbf{S}_\text{skull-pre}$ from CT scans, generating the corresponding shape latent codes $\mathbf{Z}_{shape\_face}$ and $\mathbf{Z}_{shape\_skull}$. The surgery module then fits a shared surgery latent code $\mathbf{Z}_{srg}$ by matching the deformation from $\mathbf{S}_\text{skull-pre}$ to the surgically planned postoperative skull $\mathbf{S}_\text{skull-plan}$. Finally, $\mathbf{Z}_{srg}$, $\mathbf{Z}_{shape\_face}$ and $\mathbf{F}_\text{face-pre}$ are combined in the surgery module to predict the postoperative facial surface $\mathbf{F}_\text{face-pred}$.
• Figure 2: Architecture of the shape module in NICE. Region-specific SDF decoders reconstruct facial, maxillary, and mandibular geometry from spatial coordinates and shape latent codes, using landmark-centered MLP ensembles.
• Figure 3: Architecture of the surgery module in NICE. Region-specific decoders predict deformation fields from spatial coordinates, a shared surgery latent code, and shape latent priors, enabling personalized postoperative mesh generation.
• Figure 4: Illustration of facial sub-region division for quantitative evaluation.
• Figure 5: Qualitative performance comparison of the shape and surgery modules, including ablation studies conducted against SCULPTOR-S and SCULPTOR-T. Each row corresponds to a different anatomical region: facial surface (top), maxilla (middle), and mandible (bottom). In each column, the top-right subfigure shows the P2PL error map compared with ground truth (value range: 0-5 millimeters), while the bottom-right subfigure displays magnified views of specific anatomical structures.