Program Information
Learning Dose From Anatomy
L Hibbard1*, (1) Elekta, Inc , Maryland Heights, MO
Presentations
SU-H1-GePD-J(B)-5 (Sunday, July 30, 2017) 3:00 PM - 3:30 PM Room: Joint Imaging-Therapy ePoster Lounge - B
Purpose: We are developing a deep learning (DL) model to estimate the 3D dose distribution for a novel patient CT image-set and anatomy structures, based on exemplary prior plans. Such dose estimates could aid planning and plan quality assessment.
Methods: Fifty-nine prostate cases were planned using iCycle (Erasmus MC) and Monaco (Elekta) using a constant set of prostate planning constraints and objectives. 30 cases were selected for training; 29 cases formed the ground truth (GT) for testing. For each case the anatomy was represented by three signed distance functions (SDMs; PTV, combined OARs, patient exterior). The SDMs formed the image data and the 3D dose formed the label data input to the U-Net deep convolutional neural network on the Caffe platform. The initial learning rate used is 10-2 (0-30k iterations) dropping by a factor of 0.1 each succeeding 30k increment. (10k iterations on an nVidia TitanX gpu takes about 24 hours.) DL 3D dose estimates were compared to the corresponding GT iCycle/Monaco dose distributions by examining profiles through the central PTVs and by comparing DVHs for the 3D dose distributions.
Results: PTV profiles and DVHs were computed for various DL learning regimes. DL PTV profiles are similar to the corresponding GT profiles early (5-10k) and change little after about 60k iterations. DL DVHs for all cases show less-homogeneous coverage of the PTV than the corresponding GT DVHs, with less-steep dose gradients. DL doses overall are lower than GT doses by about 5-10%. More generally, the DL model quality is sensitive to mode of data presentation, and managing sources of variation in the data is crucial.
Conclusion: DL models of plans that map anatomies to dose are challenging to construct, but will provide a reliable and detailed knowledge-based platform to guide planning and plan QA.
Contact Email: