Conditional Generative Adversarial Network for Dose Distribution Prediction In Hepatocellular Carcinoma Proton Beam Therapy: Improved Accuracy In Steep Dose Gradient Regions
Abstract
Purpose
Accurate dose prediction with deep learning technique in proton beam therapy (PBT) remains challenging due to steep dose gradients inherent to particle beams. Conventional models with pixel-wise loss functions yield over-smoothed predictions, and fixed dose gradient penalties provide only limited improvement. We developed a conditional generative adversarial network (cGAN) framework capable of dynamically penalizing high-frequency component discrepancies.
Methods
We developed a modified Pix2Pix model for three-dimensional dose prediction in PBT for hepatocellular carcinoma (HCC). The generator employed a U-Net architecture, adopting instance normalization to improve generalization. A PatchGAN discriminator enforced high-frequency detail preservation through adversarial training. A baseline U-Net with composite loss functions served as the comparison model. Both models predicted prescription-normalized dose distributions from only computed tomography (CT) images and target/organs-at-risk (OAR) structures data. Using data from 172 HCC patients, prediction accuracy was evaluated using dose-volume metrics for the clinical target volume (CTV), planning target volume (PTV), and normal liver, along with 10% isodose volume analysis representing beam path regions. Statistical significance was determined using paired t-tests.
Results
The cGAN significantly outperformed U-Net for target volume dose metrics. The mean absolute error for CTV D95% (minimum dose covering 95% of the volume) was 0.48 Gy(RBE) with cGAN versus 1.08 Gy(RBE) with U-Net (p=0.034), with similar improvements observed for D98% and D2%. PTV metrics showed trends consistent with CTV results. Normal liver metrics showed no significant differences between models (p>0.05). For the 10% isodose volume, U-Net demonstrated superior accuracy, with mean prediction errors of −7.8 cm³ compared to −38.1 cm³ for cGAN.
Conclusion
The cGAN framework significantly improved target coverage prediction by capturing steep dose gradients characteristic of PBT. Future work will focus on improving cGAN performance in the low dose region, aiming to achieve balanced and accurate predictions across all dose levels.