S Momin*, J Harms, Y Lei, S Charyyev, J Zhou, J Roper, J Bradley, T Liu, X Yang, Department of Radiation Oncology and Winship Cancer Institute, Emory Univ, Atlanta, GA

• 
  S Momin

SU-IePD-TRACK 5-3 (Sunday, 7/25/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]

Purpose: This study investigated the influence of different loss functions on dose prediction accuracy for intensity modulated proton radiation therapy(IMPT).

Methods: We proposed a dual-pyramid-network (DPN) with late-fusion-network (LFN) to generate high-quality pseudo-Monte Carlo(MC) proton plans using low-quality pencil-beam(PB)-based dose plans and corresponding CT as inputs. The proposed deep-learning network architecture includes three subnetworks: CT-only, PB dose-distribution only and late fusion subnetwork. The CT only and PB dose-distribution only subnetworks are based on U-Net architecture and extract features from the CT and PB dose distribution of each patient. Individual features are then combined in the LFN. The subnetworks were trained through three loss functions (L1-norm,L0.5-norm,and combined L1+L0.5), respectively, with constant hyperparameters in three separate models. Five-fold cross-validation experiments were performed on 40 prostate cancer patients. Predicted pseudo-MC dose distributions were compared with the target MC dose via gamma analysis under 1%/1mm, 2%/2mm and 3%/3mm acceptance criteria. Obtained dose-volume-histograms(DVHs) from each trained model were evaluated against the high-quality MC DVHs.

Results: Dose map comparison results show that L1-norm loss function performs well in low dose region, whereas L0.5-norm function performs well in high dose regions. The combined loss function achieved the best dose distribution prediction accuracy. L1norm, L0.5norm, and combined(L1+L0.5) loss functions achieved 30.8±20.3%, 71.9±17.3%, and 90.1±5.8% under 1%/1mm acceptance criteria and 56.2±26.2%, 92.5±5.6% and 95.1±0.8% under 2%/2mm acceptance criteria averaged over 40 patients from cross validation. The DVHs obtained from combined loss model had the closer agreement with the ground-truth MC DVHs compared to model with individual losses.

Conclusion: We quantitatively investigated dose prediction accuracy based on three different loss functions to generate high-quality pseudo-MC proton plans via the proposed LFN-DPN. Overall results identify the combined (L1+L0.5) loss function to be the best candidate for deep-learning-based dose prediction of IMPT and highlights the importance of choosing an appropriate loss function.

ePosters

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email