Interpretable graph-conditioned CNNs for dose-volume histogram prediction in radiotherapy

Accurate dose-volume histogram (DVH) prediction is essential for high-quality and automated radiotherapy (RT) planning. However, existing deep learning methods primarily focus on voxel-level dose prediction followed by post-processing to extract DVHs, a workflow that can introduce cumulative errors and limit clinical interpretability.

In this study, we propose a novel deep learning framework that directly predicts full DVHs for multiple organs-at-risk (OARs) from computed tomography (CT) and structure contours.

Our method combines a three-dimensional convolutional neural network (CNN) for anatomical feature extraction with a graph neural network (GNN) that models each DVH as a structured graph, enabling sequential dose-volume dependencies to be learned explicitly. The model was extensively validated on nasopharyngeal cancer cases from an external institute and rectal cancer cases from our affiliated hospital, encompassing treatments with TomoTherapy and volumetric modulated arc therapy (VMAT). We additionally conducted a model-guided clinical DVH assessment on suboptimal cases to demonstrate the method's clinical utility RESULTS: The proposed method achieved low dose-wise prediction errors and outperformed previous approach. Specifically, it reduced the mean dose error for the brain stem from 1.24 Gy to 0.66 Gy and for the larynx from 1.53 Gy to 0.78 Gy, with notable improvements also observed in the parotid glands and temporal lobes. Statistical analysis of clinical indices demonstrated non-inferiority to ground-truth plans (p > 0.05). Moreover, the predicted DVHs effectively guided plan revision and improved plan quality.

Our results suggest that direct DVH prediction using a CNN-GNN framework offers a robust and clinically interpretable solution for treatment planning and quality assurance.