Premium
TU‐C‐M100J‐06: Accurate Prediction of Intra‐Fraction Motion Using a Modified Linear Adaptive Filter
Author(s) -
Srivastava V,
Keall P,
Sawant A,
Suh Y
Publication year - 2007
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.2761337
Subject(s) - standard deviation , mathematics , gaussian , filter (signal processing) , sensitivity (control systems) , cutoff , algorithm , estimator , metric (unit) , mean squared error , statistics , computer science , physics , computer vision , operations management , quantum mechanics , electronic engineering , engineering , economics
Purpose: To predict tumor position in order to compensate for temporal latency of a target tracking‐based dynamic radiation dose delivery system. Method and Materials: A linear adaptive filter was trained for real‐time prediction of tumor target position. Signal history was first filtered by a low‐pass filter (LPF). Template matching was used in order to select training examples that were closest to the current case, using the distance metric: D j = ∑ i = 1N ( xi j − x i ) 2where,D j= distance between j th training example and test casex i j=i th component of j th training examplex i=i th component of test case .This algorithm was applied to 160 3D tumor motion datasets acquired from 46 patients. Algorithm parameters were selected through a comprehensive sensitivity analysis performed on 10% of the datasets that exhibited maximum peak‐to‐peak motion. The resulting parameters were used to analyze algorithm performance over the remaining 90% of the datasets. Root mean square (RMS), maximum, mean and standard deviation of errors with and without prediction were calculated. The error distribution was tested against a Gaussian distribution using the Kolmogorov‐Smirnov (KS) test. Results: Sensitivity analyses showed that a third‐order Butterworth LPF having cutoff frequency of 2 Hz was the most suitable. Examination of error indicates that it has close‐to‐zero mean for all the three dimensions and standard deviation is of the order of 0.1 mm. Moreover, the D‐statistics of KS test are close to 0.1 (indicating a distribution that closely approaches a Gaussian) except for cases where tumor motion itself is very small (less than 5 mm). Compared to no prediction, the algorithm reduced RMS and maximum errors, on average, by 74.58% and 55.39% respectively. Conclusion: We have developed a robust and highly efficient algorithm for predicting respiratory tumor motion. Future work includes integration of the algorithm with the laboratory setup of a dynamic MLC‐based target tracking system.