Purpose: Gating lung SBRT treatment to a limited phase range determined by 4DCT reduces the irradiated volume. Real time phase determination requires an estimate of the total cycle time prior to the actual measured end of the cycle. Breathing predictive filters provide an estimate of the total cycle length. AI with appropriately formatted data can enhance accuracy and facilitate continuous learning throughout simulation, setup, and fractionated treatments.
Methods: During 4DCT simulation 30 data samples per second are accumulated sequentially. A demonstration case with 55 cycles totaling 6738 data points was accumulated and the start of each new cycle easily identified. Cycle phase equals 100*current time/total time. All cycles are formatted into a two dimensional array consisting of the amplitude A(i,j) for cycle j at time i. The phases were divided into five ranges in one standard deviation intervals analogous to AI classification. For each range all the data were first fit to a fourth order polynomial, outliers excluded, and a polynomial fit then recomputed. The five polynomial coefficients for the five classes were stored as a 5x5 array P basically reducing the original 6738 learning data points to 25 coefficients. Additional learning data accumulated during setup and fractionated treatments are added to A and the modeling process repeated to update P.
Results: During treatment the phase for the amplitude data at each interval from the start of a new cycle is optimized using the downhill simplex method. Real time data approximately through half a cycle was compared to the five class polynomials and the error calculated. Predicted cycle length was within 5% for data within one standard deviation on the mean.
Conclusion: The proposed methodology reduces the relatively large learning data set to just 25 coefficients in five classes, is easily updated, and can predict phase in real time for accurate gated lung SBRT.
Not Applicable / None Entered.