Neural patient-specific 3D-2D registration in laparoscopic liver resection

Area of Science:

• Medical imaging
• Surgical technology
• Computer-assisted surgery

Background:

• Augmented reality (AR) in laparoscopic liver resection necessitates precise 3D-to-2D model registration.
• Challenges include liver deformation and limited intraoperative visibility.
• Current registration methods are computationally intensive and require manual initialization.

Purpose of the Study:

• To introduce the first neural network model (NM) for 3D-2D liver registration.
• To address the computational expense and manual initialization issues of existing methods.
• To enable real-time registration for enhanced surgical guidance.

Main Methods:

• A neural network model (NM) predicts 3D model deformation coefficients from 2D image landmarks.
• Patient-specific models are trained using synthetic data from preoperative 3D models.
• A liver shape modeling technique is employed to reduce computational complexity.

Main Results:

• The NM method achieves accuracy comparable to existing numerical optimization techniques.
• NM demonstrates significantly faster computation, enabling real-time inference.
• This represents a substantial advancement for intraoperative surgical guidance.

Conclusions:

• The proposed NM is the first neural network approach for 3D-2D liver registration.
• Preliminary results show competitive accuracy with superior computational efficiency.
• This method has the potential to significantly advance liver registration techniques in surgery.