Room: Exhibit Hall | Forum 6
Purpose: The goal of this study was to develop a framework for the clinical knowledge-based estimation of target position displacements (TPDs) using machine learning architectures (MLAs), i.e., artificial neural networks (ANNs) and random forests (RF), for prostate image-guided intensity-modulated radiotherapy (IG-IMRT).
Methods: The planning computed tomography (pCT) and daily cone-beam computed tomography (CBCT) images at 378 fractions for 10 patients with prostate cancer were selected for this study. The clinical target position errors between pCT and CBCT images, which were determined after an automated bone-based registration, were acquired as reference TPDs. The rectum and bladder geometrical features, which were selected from pCT and CBCT images, and corresponding reference TPDs were fed as input and teacher data, respectively, into ANNs and RF. Three different regularization theories were used in ANNs, i.e., Bayesian regularization backpropagation (BR), Levenberg-Marquardt backpropagation (LM), and scaled conjugate gradient backpropagation (SCG). TPDs were estimated by the MLAs, and the constructed MLAs were evaluated by residual errors between the reference and estimated TPDs.
Results: Since the TPDs in the left-right direction were negligible, the MLAs were constructed for estimation of TPDs along anterior-posterior (AP) and superior-inferior (SI) directions. The absolute mean residual errors in BR-ANN, LM-ANN, SCG-ANN, and RF were 0.94 mm, 0.96 mm, 0.95 mm, and 0.94 mm along the AP direction, and 0.92 mm, 0.92 mm, 0.93 mm, and 0.86 mm along the SI direction, respectively. The percentages of fractions with the residual errors less than 2 mm in BR-ANN, LM-ANN, SCG-ANN, and RF were 89.9%, 88.9%, 90.5%, and 88.9% along the AP direction, and 87.3%, 88.1%, 86.5%, and 88.4% along the SI direction, respectively.
Conclusion: The results of this study suggest that MLAs can estimate the TPDs based on clinical knowledge for patient positioning in prostate IG-IMRT.