Room: Exhibit Hall | Forum 7
Purpose: To fully automate planning for whole brain radiotherapy (WBRT). Here we examined whether a convolutional neural network (CNN) can automatically generate the field shapes based on CT images of the head.
Methods: The CT images and treatment fields of 520 clinical WBRT cases were used in this study. These cases were split into training (n=312), cross-validation (n=104), and test (n=104) sets. We first created right- and left-lateral digitally-reconstructed radiographs (DRRs) from each patientâ€™s CT scan. The DRRs and their corresponding treatment fields were used to train the deepLabV3+ architecture by only changing the weights from the last layer. Parameters involved in DRR creation, and their effect on the resulting predictions, were optimized by evaluating the trained models on the cross-validation patients. The predicted WBRT treatment fields with test set were inspected by a radiation oncologist. Quantitative analysis was then conducted on the field border. For the flat inferior field edge, the distance between the prediction and clinical field edge was measured. For points along the curved anterior-inferior field edge, the closest distance to the clinical field edge of each point was measured, and the mean distance of all points was calculated.
Results: The 104 test cases were all considered â€œvisually acceptableâ€? by the radiation oncologist. For the inferior field border, the average difference between CNN predictions and actual fields was 3.8Â±3.0 mm. Along the anterior-inferior field edge, the averaged mean prediction-to-clinical distance was 4.2Â±1.9 mm.
Conclusion: Trained with carefully selected clinical WBRT cases, the deep learning network can generate treatment fields similar to clinical fields with average field border distance ~4 mm. The predicted fields appeared consistent with the majority of training data, and they exhibited less variability than clinical fields.