[Paper] NICE: Neural Implicit Craniofacial Model for Orthognathic Surgery Prediction

Source: arXiv - 2512.05920v1

Overview

The paper introduces NICE (Neural Implicit Craniofacial Model), a deep‑learning framework that predicts how a patient’s face will look after orthognathic (jaw‑realignment) surgery. By leveraging implicit neural representations, the authors achieve higher accuracy and speed than traditional biomechanical simulators, making postoperative visualisation more reliable for surgeons and patients alike.

Key Contributions

• Implicit SDF‑based shape module: Separate neural decoders learn signed‑distance functions for the facial skin, maxilla, and mandible, enabling high‑fidelity 3‑D reconstruction from sparse clinical scans.
• Region‑specific deformation decoders: A shared surgical latent code drives dedicated decoders that output point‑wise displacement fields, capturing the nonlinear soft‑tissue response to skeletal movements.
• End‑to‑end trainable pipeline: The model learns both anatomy and surgical deformation jointly, reducing the need for hand‑crafted biomechanical parameters.
• State‑of‑the‑art performance: Quantitative benchmarks show significant error reductions in high‑impact regions (lips, chin) compared with existing parametric and deep‑learning baselines.
• Clinical‑ready inference speed: Predictive runs complete in seconds on a single GPU, suitable for intra‑operative planning tools.

Methodology

1. Data Representation – Patient CT/MRI scans are converted into point clouds. Each anatomical region (skin, maxilla, mandible) is encoded by an implicit Signed Distance Function (SDF) learned by a small multilayer perceptron (MLP). The SDF tells whether a point lies inside or outside the surface, allowing smooth, high‑resolution reconstruction without explicit meshes.

2. Shape Module – Three region‑specific SDF decoders are trained simultaneously, sharing a low‑dimensional latent vector that captures the overall craniofacial geometry of a patient.

3. Surgery Module – A surgical latent code (learned from pre‑ and post‑operative data) feeds into three deformation decoders, one per region. Each decoder predicts a 3‑D displacement vector for every input point, effectively “warping” the pre‑op anatomy to the expected post‑op shape.

4. Training Objective – The loss combines (i) SDF reconstruction error, (ii) Chamfer distance between predicted and ground‑truth post‑op point clouds, and (iii) regularizers that enforce smooth deformations and preserve anatomical constraints (e.g., bone continuity).

5. Inference – Given a new patient’s pre‑op scan and a planned skeletal movement (e.g., maxillary advancement of 5 mm), the model encodes the anatomy, injects the surgical code, and instantly outputs the predicted facial surface.

Results & Findings

• Higher fidelity in expressive regions: The model reduces errors where visual impact is greatest (lips, chin), which are traditionally hard to predict due to complex soft‑tissue dynamics.
• Anatomical consistency: Despite aggressive deformation, bone structures remain coherent, and no unrealistic self‑intersections appear.
• Robustness to limited data: Even with as few as 1,000 training cases, NICE outperforms parametric FEM approaches that require dense biomechanical meshes.

Practical Implications

• Surgical Planning Tools – Integrating NICE into existing 3‑D planning software can give surgeons instant visual feedback on proposed osteotomies, helping refine bite‑correction strategies before entering the OR.
• Patient Communication – Real‑time, high‑quality facial previews can improve informed consent and set realistic expectations, potentially reducing postoperative dissatisfaction.
• Automation of Workflow – The end‑to‑end nature eliminates manual mesh generation and finite‑element calibration, cutting down on preprocessing time and allowing clinics with limited engineering resources to adopt advanced prediction.
• Extension to Other Craniofacial Procedures – The implicit representation framework can be repurposed for trauma reconstruction, cleft palate repair, or cosmetic augmentation where soft‑tissue response matters.

Limitations & Future Work

• Dataset Diversity – The training set primarily contains adult patients of specific ethnic backgrounds; performance on pediatric or highly variable facial morphologies remains untested.
• Surgical Code Interpretability – The latent surgical vector is learned implicitly; future work could map it to explicit surgical parameters (e.g., millimeters of maxillary advancement) for better clinician control.
• Dynamic Soft‑Tissue Effects – The current model predicts static post‑op geometry; incorporating muscle activation or swelling over time would broaden its clinical relevance.
• Regulatory Pathway – While inference is fast, the authors note that extensive validation under medical device regulations is required before deployment in a clinical setting.

Authors

• Jiawen Yang
• Yihui Cao
• Xuanyu Tian
• Yuyao Zhang
• Hongjiang Wei

Paper Information

• arXiv ID: 2512.05920v1
• Categories: cs.CV, cs.LG
• Published: December 5, 2025
• PDF: Download PDF