PINNOCHIO AI speeds facial surgery planning 100x over traditional FEM
New physics-informed neural network predicts soft-tissue movement without CT ground truth.
A team led by Jungwook Lee and Pingkun Yan (among nine authors from engineering and medical schools) has released PINNOCHIO—a physics-informed neural network designed to simulate facial soft-tissue deformation for orthognathic (jaw) surgery planning. Current gold-standard finite element methods (FEM) produce highly accurate results but require hours of computation, making interactive planning impractical. Pure deep learning models, on the other hand, often violate biomechanical laws because they lack physics constraints and need full volumetric ground truth which is rarely available.
PINNOCHIO solves this with a novel hybrid sequential decomposition: it first predicts the discontinuous movement at the bone–soft-tissue interface, then simulates continuous hyperelastic deformation of the volume. This separation allows the network to train stably using only surface scans (e.g., from stereophotogrammetry) as supervision, while the physics-informed losses enforce internal consistency. On a 40-patient cohort, PINNOCHIO surpasses baseline deep learning and FEM-hybrid methods in both surface accuracy and physical plausibility. Crucially, it runs orders of magnitude faster than FEM, enabling real-time interactive feedback for surgeons to explore surgical alternatives within a single consultation.
- PINNOCHIO uses a two-stage decomposition: first predicts bone–soft-tissue interface movement, then volumetric hyperelastic deformation.
- Trained on only outer facial surface data (no internal CT labels) by enforcing physics constraints, achieving biomechanical consistency.
- Evaluation on 40 patients shows higher surface accuracy and physical validity than existing baselines, with dramatic speedup over FEM.
Why It Matters
Surgeons can iterate orthognathic surgery plans in minutes instead of hours, improving outcomes without costly simulations.