Predicting the outcome of respiratory motion prediction.

PURPOSE: Prediction of respiratory motion traces has become an important research topic. Especially for motion compensated radiotherapy, compensation of the latencies arising from mechanical constraints and signal processing is necessary. In recent years, many algorithms have been developed and evaluated. It is, however, still unclear how well a specific patient will be suited to motion prediction before the treatment actually starts.
METHODS: In this work, we have analyzed 304 respiratory motion traces with an average duration of 71 min. A total of 21 features characterizing these signals (12 from the frequency domain and 9 from the time domain) have been determined for each motion trace. The correlation between these features and the overall prediction quality for three different algorithms (based on wavelet-based multiscale autoregression, support vector regression, and linear expansion of the prediction error) has been analyzed and six dominant features have been identified (three each from the time and frequency domains). Additionally, the optimized results of the multistep-linear method (MULIN) prediction algorithm on the first 300 s of motion data have been used as a seventh, independent feature. Assessing the prediction algorithms' quality was done by calculating the relative root mean squared (RMS(rel)) error, i.e., the ratio between the RMS error of the prediction output and the RMS error of the delayed signal (the RMS error obtained when doing no prediction). Then, for each algorithm, the signals themselves were grouped into four classes according to the quality of prediction: relative RMS less than 0.8 (C1), between 0.8 and 0.9 (C2), between 0.9 and 1.0 (C3), and over 1.0 (C4). The goal of this work is to identify, prior to treatment, those patients whose respiratory behavior indicates probable (RMS(rel) ≥ 0.9) or certain (RMS(rel) ≥ 1.0) failure of respiratory motion prediction. Consequently, all signals from C4 must be identified and rejected and no signals from C1 may be falsely rejected. The restriction on C2 and C3 is slightly weaker: C2 are those signals that should be kept and C3 are those signals that should be rejected.
RESULTS: Rejecting all signals from C4 and C3, keeping as many signals from C1 and as few from C2 as possible, has been achieved for the wLMS algorithm when using six feature pairs and the result of prediction on the short signal. Here, the false rejectance rate for C1 was less than 13% and the false acceptance rate for C2 was 15%. For the SVRpred and MULIN algorithms, the results are somewhat worse: in both cases, signals from C3 were falsely accepted (25.0% and 14.3%, respectively) but all signals from C4 were rejected. The false rejectance rate for C1 was 11.4% (MULIN) and 26.3% (SVRpred).
CONCLUSIONS: In general, it has been shown that pretreatment classification of the quality of respiratory motion prediction is possible and that signals with high relative RMS error can be identified with great reliability. This is especially true for the wLMS algorithm, which has also been identified as the most precise and robust of the presented methods.