Efficient estimation of tissue thicknesses using sparse approximation for Gaussian processes | JoVE Visualize

Area of Science:

Medical Physics
Biomedical Engineering
Optical Imaging

Background:

Accurate human skull localization is crucial for effective cranial radiotherapy.
Marker-less optical head tracking offers a rapid and precise method for monitoring patient motion during treatment.
Incorporating tissue thickness, alongside 3D surface geometry, enhances marker-less forehead tracking accuracy.

Purpose of the Study:

To investigate sparse approximation methods for Gaussian Processes (GPs) to reduce computational complexity in tissue thickness estimation.
To evaluate the trade-off between computational time and root mean square error (RMSE) for different sparse approximation techniques.
To assess the efficacy of data clustering prior to subset selection for improving GP performance.

Main Methods:

Applied Gaussian Processes (GPs) utilizing optical backscatter from a sweeping laser to determine tissue thickness.
Evaluated sparse approximation techniques, including a subset of data (SoD) method and inducing point methods (FITC), across five subjects.
Compared computational time and RMSE of full GP models against sparse approximations.
Investigated the impact of clustering training data before applying subset selection.

Main Results:

The subset of data (SoD) technique demonstrated a favorable trade-off between computation time and RMSE compared to other methods.
The increase in RMSE when reducing data points with SoD was not significant enough to warrant the computational overhead of FITC.
Clustering the training data before subset selection yielded promising results, suggesting improved efficiency and accuracy.

Conclusions:

Sparse approximation, particularly the SoD technique, is effective in reducing the computational burden of GP-based tissue thickness estimation for cranial radiotherapy.
Clustering training data prior to subset selection shows potential for further optimizing marker-less optical head tracking systems.
These findings contribute to developing more efficient and accurate real-time motion monitoring in radiotherapy.