CT-based Unsupervised DL for orthognathic surgery

Unsupervised learning of reference bony shapes for orthognathic surgical planning with a surface deformation network

Authors: Deqiang Xiao, Hannah Deng, Chunfeng Lian, Tianshu Kuang, Qin Liu, Lei Ma, Yankun Lang, Xu Chen, Daeseung Kim, Jaime Gateno, Steve Guofang Shen, Dinggang Shen, Pew-Thian Yap, James J. Xia

Summary:

This study proposes an unsupervised geometric deep learning framework to generate patient-specific reference bony shape models for orthognathic surgical planning. Using CT Scan from 47 normal subjects, 61 patients with jaw deformities (training); 24 paired pre and postoperative patients (testing). A Surface Deformation Network (SDNet) operating on 3D point clouds is introduced to warp a patient’s deformed craniomaxillofacial bone toward a dictionary of normal bone shapes. Sparse representation learning is then applied to estimate an objective reference model without requiring paired pre and postoperative data or extensive manual landmark digitization. The proposed method significantly outperforms landmark based sparse representation while preserving midface anatomy.

Extracted Data:

Year: 2021

Imaging Modality: CT

Dataset: 47 normal subjects, 61 patients with jaw deformities (training); 24 paired pre and postoperative patients (testing)

Dataset Split:

• Training: 47 normal subjects and 61 patients with jaw deformities (used to generate 2867 random pairs).

• Testing: 24 patients (paired pre- and postoperative data)

ML Task: 3D shape estimation / surface deformation (unsupervised learning)

Network Architecture: Surface Deformation Network (SDNet) based on PointNet++ and PointConv

Metrics: Vertex Distance (VD), Edge Distance (ED), Surface Coverage (SC), Landmark Distance (LD)

AP - Professional Qty: 1 (Landmarks were localized by "an experienced oral surgeon")

AP - Supervisor Presence: No info

AP - Experience Level: Unavailable

AP - Expertise Area: OMFS

AP - Tool or System: Unavailable

ML Task Unsupervised surface deformation / 3D shape estimation (predicting vertex-wise displacements)

Project Objective To reduce the experience dependence in orthognathic surgical planning by generating objective, patient-specific reference facial bone shape models

Clinical Relevance:

This framework reduces surgeon experience dependency during orthognathic surgical planning by providing an objective, patient‑specific reference bone model. The method improves planning reproducibility and accuracy while minimizing reliance on manual landmark placement, with potential benefits for surgical outcomes and workflow standardization.

Why is this study classified as Unsupervised Learning?

Although this work employs a deep neural network, it is considered unsupervised learning because no paired normal–deformed ground truth exists in clinical practice. The model is trained using unpaired CT scans of normal subjects and patients with jaw deformities, learning anatomical normality through geometric constraints and surface deformation rather than explicit target labels.