Room: Stars at Night Ballroom 1
Purpose: To compare deep learning architectures for dose prediction in knowledge-based automated planning (KBAP).
Methods: We trained three knowledge-based planning (KBP) models with the same 130 clinical oropharyngeal treatment plans to predict dose distributions from contoured CT images. Two of the models used generative adversarial networks (GANs) to predict dose as (1) single 2D-axial slices (2D-GAN) and (2) full 3D patient volumes (3D-GAN). The third model was based on DoseNet, a previously published state-of-the-art approach using a 3D convolutional neural network. Each model was used to predict the dose distribution for 87 out-of-sample oropharyngeal patients. Dose predictions were then input into a conventional optimization model to generate fluence-based treatment plans. Our plans were benchmarked against the corresponding clinical plans using clinical planning criteria to quantify plan quality and the gamma index (3%/3mm) to quantify spatial similarity.
Results: 3D-GAN plans satisfied the same planning criteria as the clinical plans more frequently (78%) than the 2D-GAN (76%) and DoseNet (72%) plans. Both the 3D-GAN and 2D-GAN plans satisfied 100% of the primary target volume criteria, which was 10 percentage points more than the DoseNet plans. On average, the DoseNet and 3D-GAN plans also achieved criteria for the parotids by a margin of 0.1Gy and 1.0Gy better than the 2D-GAN and clinical plans, respectively. However, on average, the DoseNet plans passed the gamma criteria 2% and 5% more frequently than the 2D-GAN and 3D-GAN plans, respectively.
Conclusion: 3D-GAN predictions in our KBAP pipeline produced the highest quality plans, followed by the 2D-GAN model and finally DoseNet (i.e., a 3D-model). The 3D-GAN appears to be better suited for KBAP over a complex patient cohort than 2D-GAN and DoseNet, and we hypothesize it is because mapping a CT image to a dose prediction is akin to style-transfer problems that GANs are particularly well suited for.
Not Applicable / None Entered.
Not Applicable / None Entered.