A prediction model for dosimetric-based lung adaptive radiotherapy

Anatomical changes occurred during the treatment course of radiation therapy for lung cancer patients may introduce clinically unacceptable dosimetric deviations from the planned dose. Adaptive radiotherapy (ART) can compensate these dosimetric deviations in subsequent treatments via plan adaption. Determining whether and when to trigger plan adaption during the treatment course is essential to the effectiveness and efficiency of ART. In this study, we aimed to develop a prediction model as an auxiliary decision-making tool for lung ART to identify the patients with intrathoracic anatomical changes that would potentially benefit from the plan adaptions during the treatment course.

Seventy-one pairs of weekly cone-beam computer tomography (CBCT) and planning CT (pCT) from 17 advanced non-small cell lung cancer patients were enrolled in this study. To assess the dosimetric impacts brought by anatomical changes observed on each CBCT, dose distribution of the original treatment plan on the CBCT anatomy was calculated on a virtual CT generated by deforming the corresponding pCT to the CBCT and compared to that of the original plan. A replan was deemed needed for the CBCT anatomy once the recalculated dose distribution violated our dosimetric-based trigger criteria. A three-dimensional region of significant anatomical changes (region of interest, ROI) between each CBCT and the corresponding pCT was identified, and 16 morphological features of the ROI were extracted. Additionally, eight features from the overlapped volume histograms (OVHs) of patient anatomy were extracted for each patient to characterize the patient-specific anatomy. Based on the 24 extracted features and the evaluated replanning needs of the pCT-CBCT pairs, a nonlinear supporting vector machine was used to build a prediction model to identify the anatomical changes on CBCTs that would trigger plan adaptions. The most relevant features were selected using the sequential backward selection (SBS) algorithm and a shuffling-and-splitting validation scheme was used for model evaluation.

Fifty-five CBCT-pCT pairs were identified of having an ROI, among which 21 CBCT anatomies required plan adaptions. For these 21 positive cases, statistically significant improvements in the sparing of lung, esophagus and spinal cord were achieved by plan adaptions. A high model performance of 0.929 AUC (area under curve) and 0.851 accuracy was achieved with six selected features, including five ROI shape features and one OVH feature. Without involving the OVH features in the feature selection process, the mean AUC and accuracy of the model significantly decreased to 0.826 and 0.779, respectively. Further investigation showed that poor prediction performance with AUC of 0.76 was achieved by the univariate model in solving this binary classification task.

We built a prediction model based on the features of patient anatomy and the anatomical changes captured by on-treatment CBCT imaging to trigger plan adaption for lung cancer patients. This model effectively associated the anatomical changes with the dosimetric impacts for lung ART. This model can be a promising tool to assist the clinicians in making decisions for plan adaptions during the treatment courses.