Randomly Wired Neural Networks for IMRT 3D dose prediction in oropharyngeal cancer

Abstract

Intensity-modulated radiation therapy (IMRT) requires patient-specific three-dimensional dose distributions that maximize target coverage while limiting irradiation of surrounding organs at risk (OARs). Although deep-learning approaches can accelerate dose prediction, commonly used architectures such as U-Nets rely on manually designed connectivity patterns that constrain exploration of the architectural search space. This study investigates randomly wired neural networks (RWNNs) for oropharyngeal cancer dose prediction. Architectures were generated from three stochastic graph families—Erdős-Rényi, Barabási-Albert, and Watts-Strogatz—and optimized using Bayesian optimization. Performance was evaluated using five-fold cross-validation and five random seeds against a U-Net baseline across voxel-wise, spatial, target-volume, and OAR metrics. RWNNs achieved comparable or significantly improved performance across several metrics, including dose error, spatial agreement, PTV70D95, homogeneity, and spinal-cord Dmax, although the U-Net retained superior brainstem dose accuracy. These results were achieved with substantially fewer trainable parameters, with the Watts-Strogatz architecture containing nearly 30-fold fewer parameters than the U-Net, although RWNNs required greater activation memory and inference time. Seed analysis indicated generally stable performance, while node ablation revealed no significant deterioration after removal of individual nodes, suggesting architectural redundancy and opportunities for further compression. These findings demonstrate that stochastic graph-based architecture search can identify compact networks capable of competitive IMRT dose prediction. Determining how graph properties such as connectivity, clustering, path length, and hub structure influence dose-prediction performance may provide a mechanistic basis for designing task-specific neural architectures.