Continuous Neural ODE–Based Reconstruction for Ultra-Sparse-View CBCT In Radiotherapy
Abstract
Purpose
Ultra-sparse-view cone-beam CT (CBCT) acquisition reduces imaging dose in image-guided radiotherapy; however, extreme angular sparsity introduces severe streak artifacts and structural inconsistency that challenge conventional and learning-based reconstruction methods. Most deep learning approaches rely on discrete slice-wise mappings, lacking a continuous formulation and may be unstable under ultra-sparse conditions. This study proposes a Neural Ordinary Differential Equation (Neural ODE)–based framework that formulates CBCT reconstruction as a continuous transformation in latent space.
Methods
A dataset of 633 head-and-neck planning CT scans was used. Ultra-sparse-view CBCT data were generated via physics-based forward projection, sparse angular sampling, and FDK reconstruction at three sampling rates (1/64, 1/32, 1/16 of full views). UNODER integrates an encoder, a Neural ODE latent dynamics module, and a decoder to model reconstruction as a continuous latent flow from CBCT to CT. This enforces smoothness and consistency, improving robustness to extreme sparsity. Quantitative evaluation used PSNR and SSIM, compared with state-of-the-art deep learning baselines. Dose recalculation and gamma analysis (1 mm/1 %) assessed physical relevance.
Results
Across all sampling rates, UNODER showed superior reconstruction compared with competing methods. Under the most severe sparsity (1/64), it achieved mean PSNR 32.40 dB and SSIM 0.924, with PSNR gains up to 1.6 dB over baselines. The method enhanced streak suppression and inter-slice structural consistency. Dose recalculation showed high gamma passing rates relative to reference CTs, with mean Global gamma 96.75%, indicating preserved physical fidelity.
Conclusion
Continuous Neural ODE formulation in UNODER provides a physically motivated, robust solution for ultra-sparse-view CBCT reconstruction, supporting low-dose image-guided and adaptive radiotherapy.