(Sunday, 7/14/2019) 2:00 PM - 3:00 PM
Room: Stars at Night Ballroom 2-3
Purpose: Convolutional neural networks (CNNs) are gaining interest to predict 3D dose distribution for automatic radiotherapy treatment planning. While recent publications mainly focused on dose prediction from contours (Nguyen et al, 2018), we investigated the feasibility to predict the dose distribution from the planning CT directly without prior delineation. In the future, multitask learning combining delineation and dose prediction may result in a more efficient workflow for adaptive radiotherapy.
Methods: Our dataset consisted of 65 prostate cancer patients treated with volumetric modulated arc therapy (VMAT) with a prescription dose of 77Gy, of which 28 received a focal boost to the gross tumor volume (GTV). A patch-based regressional CNN was trained to predict the 3D dose distribution directly from the patient’s CT image. In case the patient received a boost, the GTV delineation was provided as extra input. In addition, also the position of the isocenter and the contours used for creating the ground truth dose maps could be provided as extra input. To evaluate the feasibility of dose prediction from CT and to assess the effect of additional spatial and contour information, dose predictions from different CNNs without isocenter or contour information (No ISO), with isocenter information (ISO), and with isocenter and contour information (ISO+Contours) were evaluated using dose volume histograms (DVH), dose profile curves (DPC) and the maximal dose (Dmax).
Results: Our results indicate that direct dose prediction from CT images is feasible, especially when isocenter position information is included. Adding contours as input further improves the dose distributions according to DVH, DPC and Dmax.
Conclusion: We demonstrate the feasibility of predicting dose distributions directly from CT images using additional isocenter position information. However, contour information remains important and should be included in the network. This indicates the potential of combined delineation and dose prediction approaches.
Funding Support, Disclosures, and Conflict of Interest: Siri Willems is supported by a Ph.D. fellowship of the research foundation, Flanders (FWO).
Keywords
Taxonomy
IM/TH- Image Analysis (Single modality or Multi-modality): Machine learning
Contact Email
