Purpose: To quantify anatomical changes of lung cancer patients and evaluate the resulting dosimetric effect according to on-treatment cone-beam computed tomography (CBCT) scans. To develop a dosimetric-based lung adaptive radiotherapy (ART) decision tool based on features of patients’ anatomy and anatomical changes.
Methods: Eleven non-small cell lung cancer patients were enrolled in this study. A planning CT (pCT) and weekly CBCTs of each patient were collected. A virtual CT (vCT) was generated by deforming the pCT to each CBCT and the dose distribution was recalculated on the vCT (DvCT). The replanning status of a CBCT was labeled as positive if the DvCT was clinically unacceptable or could be potentially improved according to two sets of dosimetric criteria, otherwise labeled as negative. The region of anatomical changes (region of interest, ROI) between each CBCT and the pCT was extracted. Twenty-four features from the ROI and the pCT anatomy were extracted for each pCT-CBCT pair. A model based on nonlinear supporting vector machine was built for replanning status prediction. A nested cross-validation (CV) scheme was used for model evaluation, where the recursive feature elimination method was applied for feature selection.
Results: Thirty-Five CBCT-pCT pairs were identified of having a ROI, among which only 7 CBCTs had positive replanning status. Six features were selected for model fitting and the top 6 most frequently selected features were from both feature categories. A high prediction accuracy of 0.84 was achieved and the area under the mean curve was 0.82. A balanced sensitivity and specificity of 0.70 could be obtained by lowering the probability threshold.
Conclusion: Our preliminary results demonstrated that the proposed model has the potential to identify the replanning need for lung cancer patients from CBCTs. Further study will be done to alleviate the imbalanced class problem and refine the model performance.
Not Applicable / None Entered.