Kernel density estimation-based real-time prediction for respiratory motion.

Effective delivery of adaptive radiotherapy requires locating the target with high precision in real time. System latency caused by data acquisition, streaming, processing and delivery control necessitates prediction. Prediction is particularly challenging for highly mobile targets such as thoracic and abdominal tumors undergoing respiration-induced motion. The complexity of the respiratory motion makes it difficult to build and justify explicit models. In this study, we honor the intrinsic uncertainties in respiratory motion and propose a statistical treatment of the prediction problem. Instead of asking for a deterministic covariate-response map and a unique estimate value for future target position, we aim to obtain a distribution of the future target position (response variable) conditioned on the observed historical sample values (covariate variable). The key idea is to estimate the joint probability distribution (pdf) of the covariate and response variables using an efficient kernel density estimation method. Then, the problem of identifying the distribution of the future target position reduces to identifying the section in the joint pdf based on the observed covariate. Subsequently, estimators are derived based on this estimated conditional distribution. This probabilistic perspective has some distinctive advantages over existing deterministic schemes: (1) it is compatible with potentially inconsistent training samples, i.e., when close covariate variables correspond to dramatically different response values; (2) it is not restricted by any prior structural assumption on the map between the covariate and the response; (3) the two-stage setup allows much freedom in choosing statistical estimates and provides a full nonparametric description of the uncertainty for the resulting estimate. We evaluated the prediction performance on ten patient RPM traces, using the root mean squared difference between the prediction and the observed value normalized by the standard deviation of the observed data as the error metric. Furthermore, we compared the proposed method with two benchmark methods: most recent sample and an adaptive linear filter. The kernel density estimation-based prediction results demonstrate universally significant improvement over the alternatives and are especially valuable for long lookahead time, when the alternative methods fail to produce useful predictions.