3D craniofacial generative model for surgical planning in mandibular reconstruction

Area of Science:

• Medical Imaging
• Computer-Aided Surgery
• Artificial Intelligence in Medicine

Background:

• Mandibular reconstruction after oral tumor resection is challenging, requiring restoration of function and aesthetics.
• Current Computer-Assisted Surgery (CAS) workflows are fragmented, using manual mirroring and labor-intensive planning.
• Existing deep learning models lack integration of surgical feasibility and soft-tissue outcome prediction.

Purpose of the Study:

• To develop a unified craniofacial generative framework for integrated mandibular reconstruction.
• To automate mandibular shape completion, surgical planning, and postoperative facial prediction.
• To enhance both functional and aesthetic outcomes in mandibular reconstruction.

Main Methods:

• Utilized a 3D latent diffusion model with patch-wise encoding for anatomical shape prior learning.
• Employed a specialized encoder for high-fidelity completion of defective mandibles.
• Introduced a dynamic programming algorithm for automated fibula osteotomy and splicing planning.
• Integrated postoperative facial morphology prediction conditioned on the reconstructed bone.

Main Results:

• Achieved high fidelity in mandibular completion (Dice 85.61%, CD 1.43 mm) and facial prediction (Dice 97.67%, CD 1.57 mm).
• Improved reconstruction precision with reduced volume ratio (28.28%), contour error (2.24 mm), and max projection (3.55 mm).
• Reduced total planning time from over 34 minutes to under one minute (60x speedup).

Conclusions:

• The proposed unified framework offers a practical and efficient solution for aesthetically aware mandibular reconstruction.
• The model integrates shape completion, surgical planning, and outcome prediction within a single pipeline.
• This approach supports improved functional and aesthetic restoration in reconstructive surgery.