Data assimilation using a gradient descent method for estimation of intraoperative brain deformation.

Area of Science:

• Computational Biomechanics
• Medical Image Analysis
• Surgical Navigation

Background:

• Biomechanical models are crucial for brain shift compensation during surgery.
• Current methods like the forced-displacement method have limitations due to measurement uncertainty and fictitious forces.
• The Representer algorithm (REP) improves accuracy but is computationally inefficient with increasing data points.

Purpose of the Study:

• To develop a computationally efficient algorithm for brain deformation estimation.
• To minimize the difference between measured and model-estimated displacements using sparse data.
• To enable accurate and fast brain shift compensation in real-time surgical settings.

Main Methods:

• Introduction of a steepest gradient descent (SGD) algorithm for biomechanical modeling.
• Iterative adjustment of forcing conditions to minimize model-data misfit.
• Utilizing a parallelized direct solver on a shared-memory Linux cluster for solving linear systems.

Main Results:

• SGD achieves comparable or superior model estimates to REP, capturing 74-82% of tumor displacement.
• SGD demonstrates a significant computational effort reduction (4-fold or more) compared to REP.
• Computational cost is nearly invariant to the number of assimilated data points, with ~2 min for 100 points in patient cases.

Conclusions:

• The SGD algorithm provides a computationally efficient and accurate solution for estimating whole-brain deformation.
• SGD's speed and accuracy make it suitable for routine use in the operating room for brain shift compensation.
• This advancement has the potential to improve surgical navigation and patient outcomes.