Online prediction of respiratory motion: multidimensional processing with low-dimensional feature learning

Area of Science:

• Medical Physics
• Radiotherapy
• Image-guided therapy

Background:

• Accurate real-time prediction of respiratory motion is crucial for effective radiotherapy targeting lung tumors.
• Existing 1D nonparametric methods show promise, but extending them to multidimensional prediction faces the 'curse of dimensionality' due to exponential data requirements.
• Multidimensional prediction requires handling high-dimensional data with limited training instances.

Purpose of the Study:

• To develop and evaluate a multidimensional prediction scheme for respiratory motion using kernel density estimation (KDE) in an augmented space.
• To address the 'curse of dimensionality' by utilizing principal component analysis (PCA) for feature extraction and constructing a low-dimensional manifold.
• To improve the accuracy of real-time respiratory motion prediction for lung cancer radiotherapy.

Main Methods:

• Investigated a multidimensional prediction scheme based on kernel density estimation (KDE).
• Utilized principal component analysis (PCA) to construct a low-dimensional feature space, alleviating the 'curse of dimensionality'.
• Compared the proposed low-dimensional feature learning method against independent prediction along each physical coordinate using 159 lung target motion traces.

Main Results:

• The proposed PCA-based multidimensional prediction method demonstrated uniformly better performance than independent prediction.
• Reduced the case-wise 3D root mean squared prediction error by approximately 30-40%.
• Significantly improved prediction accuracy, reducing the 90% percentile 3D error from 1.80 mm to 1.08 mm (160 ms) and 2.76 mm to 2.01 mm (570 ms).

Conclusions:

• The developed method effectively handles high-dimensional respiratory motion data with limited training samples.
• The dimension reduction technique reveals insights into motion dynamics, separating semiperiodic components from random noise for prediction.
• This approach offers a pathway for improved motion management in radiotherapy and has broader applications in high-dimensional data processing.