Dosimetric Uncertainties in Dominant Intraprostatic Lesion Simultaneous Boost Using Intensity Modulated Proton Therapy

Abstract

Purpose: While intensity modulated proton therapy can deliver simultaneous integrated boost (SIB) to the dominant intraprostatic lesion (DIL) with high precision, it is sensitive to anatomic changes. We investigated the dosimetric effects from these changes based on pretreatment cone-beam computed tomographic (CBCT) images and identified the most important factors using a multilayer perceptron neural network (MLPNN). Methods and Materials: DILs were contoured based on coregistered multiparametric magnetic resonance images for 25 previously treated prostate cancer patients. SIB plans were created with (1) prostate clinical target volume − V70 Gy = 98%; (2) DIL − V98 Gy > 95%; and (3) all organs at risk within clinical constraints. SIB plans were applied to daily CBCT-based deformed planning computed tomography (CT). DIL − V98 Gy, bladder/rectum maximum dose (Dmax) and volume changes, femur shifts, and the distance from DIL to organs at risk in both planning computed tomograms and CBCT were calculated. Wilcoxon signed-ranks tests were used to compare the changes. MLPNNs were used to model the change in ΔDIL − V98 Gy > 10% and bladder/rectum Dmax > 80 Gy, and the relative importance factors for the model were provided. The performances of the models were evaluated with receiver operating characteristic curves. Results: Comparing initial plan to the average from evaluation plans, respectively, DIL − V98 Gy was 89.3% ± 19.9% versus 86.2% ± 21.3% (P = .151); bladder Dmax 71.9 ± 0.6 Gy versus 74.5 ± 2.9 Gy (P 10%, DIL to rectum distance changes, DIL to bladder distance changes, and rectum volume changes ratio are the 3 most important factors. The areas under the receiver operating characteristic curves were 0.89, 1.00, and 0.99 for the modeling of ΔDIL − V98 Gy > 10%, and bladder and rectum Dmax > 80 Gy, respectively. Conclusions: Dosimetric changes in DIL SIB with intensity modulated proton therapy can be modeled and classified based on anatomic changes on pretreatment images by an MLPNN.