NICE: Neural Implicit Craniofacial Model for Orthognathic Surgery Prediction

This is the official code repository of our work "NICE: Neural Implicit Craniofacial Model for Orthognathic Surgery Prediction" accepted by AAAI 2026.

Installation

You can create an anaconda environment called NICE using

conda create -n "NICE" python=3.9
conda activate NICE

To install this repository and its requirements in development mode run

pip install -e .

If you plan to run a model you will need have a GPU-enabled pytorch version running. For our experiemnts we used pytorch 1.13 and CUDA 11.6.

Data structure

Shape Data

dataset used for the shape module is organized into separate directories for training, validation, and testing. Each folder is numerically labeled (e.g., 000) to represent an individual sample. For each sample, the following files are included:

• softtissue.obj, maxilla.obj and mandible.obj: These files contain the ground truth (GT) data of the soft tissue, maxilla and mandible, respectively, which are obtained through CT scans.

• sft_x_face.npy, mxl_x_face.npy, and mdb_x_face.npy: These files correspond to the front surface points sampled from the soft tissue and the maxilla and mandible, where 'x' represents the sample index.

• sft_x_non_face.npy and mxl_x_non_face.npy: These files contain the back surface points sampled from the soft tissue and the maxilla, where 'x' represents the sample index.

Surgery Data

surgerydataset used for the surgery module is organized into separate directories for training, validation, and testing. Each folder is numerically labeled (e.g., 000) to represent an individual sample. For each sample, the following files are included:

• ori_sft_pre.obj, ori_mxl_pre.obj and ori_mdb_pre.obj: These files contain the preoperative ground truth (GT) data of the soft tissue, maxilla and mandible, respectively, which are obtained through CT scans.
• ori_sft_post.obj, ori_mxl_post.obj and ori_mdb_post.obj: These files contain the postoperative ground truth (GT) data of the soft tissue, maxilla and mandible, respectively, which are obtained through CT scans.
• sft_corresp_x.npy, mxl_corresp_x.npy, and mdb_corresp_x.npy: These files contain the preoperative and postoperative coordinates of sampling points on the soft tissue, maxilla, and mandible, respectively, where 'x' represents the sample index. These points share the same topology, allowing for accurate comparison of the anatomical structures before and after the surgical procedure.
• sft_corresp_normals_x.npy, mxl_corresp_normals_x.npy, and mdb_corresp_normals_x.npy: These files contain the preoperative and postoperative normal vectors of sampling points on the soft tissue, maxilla, and mandible, respectively, where 'x' represents the sample index. These points share the same topology, allowing for accurate comparison of the anatomical structures before and after the surgical procedure.

Training

NICE is trained in two stages.

Stage 1 - Learning Shape Module

First you need to train the shape module using:

python scripts/train/train_shape.py -cfg_file scripts/config/NICE_shape.yaml -local -exp_name SHAPE_EXP_NAME 

Stage 2 - Learning Surgery Module

Afterwards, the surgery module can be trained by:

python scripts/train/train_surgery.py -cfg_file scripts/config/NICE_surgery.yaml -exp_name SURGERY_EXP_NAME -mode compress

Test

Test Shape Module

You can test the shape module using

python scripts/fit/fit_shape.py -cfg_file scripts/config/NICE_shape.yaml -local 

Test Surgery Module

You can test the surgery module using

python scripts/fit/fit_surgery.-cfg_file scripts/config/NICE_surgery.yaml -exp_name SURGERY_FIT_NAME -mode compress

Predict

You can predict the postoperative face using

python scripts/fit/predict_face.py -cfg_file scripts/config/NICE_surgery.yaml -exp_name SURGERY_PREDICT_NAME -mode compress

Data Availability

Due to the data confidentiality agreement with the hospital, we can only provide a few demo cases. The complete dataset will be made publicly available in the future.