Optimizing material composition determination in dual-energy computed tomography: a comparative study of a linear model and a fully connected neural network

Accurate elemental decomposition in dual-energy computed tomography (DECT) is crucial for precision in radiation therapy planning. We present a comparative study of linear regression and fully connected neural networks (FCNNs) for voxel-wise prediction of tissue elemental composition, using synthetic datasets that incorporate realistic intra- and inter-patient variability. Both models performed well under noise-free conditions, with linear regression yielding slightly lower errors. Under noisy conditions, performance degraded for both models, though the linear model generally retained lower numerical error. The FCNNs, however, consistently produced physically plausible (non-negative) elemental mass-fraction estimates. These models are well suited for integration into model-based iterative reconstruction algorithms to support artificial intelligence-driven radiation treatment planning. Future work should incorporate elemental covariances and spatial context to enhance accuracy and clinical utility.