Decision fusion for temporal prediction of respiratory liver motion

In this study we propose to combine the outcome of four temporal prediction methods, which have different strength and weaknesses, by taking their median. Based on 25 motion traces from ultrasound liver tracking, this decision fusion provided statistically significantly better results than the individual outcomes for latencies from 150 to 1000ms. On average, RMS errors reduced by at least 50% in comparison to assuming no motion for all latencies. Furthermore it was shown that time-intensive optimization of the methods parameters to individual cases was not required, as performance from using the median parameters from the population was not significantly worse when using decision fusion.