A Kernel Method for Real-Time Respiratory Tumor Motion Estimation Using External Surrogates

Abstract

Respiratory tumor motion is a major challenge in radiation therapy. Real-time tumor tracking aims to dynamically target the tumor with the radiation beam, allowing a reduction in the volume of healthy tissue exposed to a high dose. Fluoroscopic tracking with implanted fiducial markers is not widely accepted due to its invasiveness and, to a lesser extent, additional imaging dose. Using external respiratory surrogates is an attractive approach to localizing tumors affected by respiratory motion. In this paper, we present a kernel regression approach to the respiratory tumor motion problem. Due to the local nature of the kernel functions, it is inherently robust to outliers in the training data. In addition, both training and adapting to new data is highly efficient and almost free under the kernel regression model. We evaluated the method on respiratory motion data from six patients (three lung and three pancreas patients). It was found that the average 3D error is < 1 mm; the 95th percentile error is < 3 mm (1.2 mm on average). These errors are well below the peak-to-peak amplitude (10 mm on average) for all six patients and within clinically acceptable ranges. Overall, the method is able to capture highly nonlinear and complex relations between the internal tumor motion and external surrogates in an accurate and efficient (~1 ms per prediction) way. These nice properties make it a suitable candidate for accurate and robust tumor tracking using external respiratory surrogates.