Purpose: Create a 3D deep learning model capable of generating a Planning Target Volume (PTV) on the prostate and nodes volume with CT input images for radiotherapy treatment planning.
Methods: We anonymized and exported 265 CT images and PTV_4500 volumes of prostate cancer patients for this project. PTVs with 45 Gy were contoured by 7 MDs and reviewed to be clinically acceptable. Two groups were formed from 212 cases for training and 53 for testing. We developed a Pixel-Deconvolutional Network (PDN) model by tweaking the weighting parameters to minimize the dice loss function using the training cases. The performance of the auto-segmentation was analyzed through the similarity DICE and Average Hausdorff Distance (AHD) between manual (gold-standard) and auto segmented PTV_4500 (predicted). Given the restrictions of computing resources, all images required resizing to lower resolution initially as input for the network.
Results: The mean DICE/AHD scores with standard deviation (SD) for the testing set were (0.716±0.076) and (2.061±1.749 cm) for the CT image modality. If excluding one testing case with AHD > 10 cm, the mean and SD were (0.723±0.062) and (1.885±1.209 cm). The computation time for training was about 222 minutes.
Conclusion: A deep-learning network was developed to contour the PTV for radiotherapy of prostate with nodes to 45 Gy with simulation CT images. The predicted contours have a reasonable dice score and AHD value compared with the manual contouring. We believe increased contouring accuracy is possible through more curating of the training data.