Task-Based Evaluation of Sparse-View CT Reconstruction Metrics for Radiotherapy Imaging
Abstract
Purpose
To test whether commonly used pixel-wise CT reconstruction metrics reflect preservation of clinically relevant anatomy for radiotherapy imaging, and to develop an anatomy-centered, task-based evaluation and enhancement approach for sparse-view reconstruction.
Methods
We defined an anatomy-centered “completeness” evaluation suite spanning large organs (N=5), small organs (N=5), vascular structures (N=5), and pancreatic tumors (N=3). Structural preservation was measured using segmentation-derived metrics, including normalized surface Dice (NSD) for organs and clDice for tubular vasculature, to quantify boundary fidelity and connectivity that are important for planning and QA. We then developed Completeness-Aware Reconstruction Enhancement (CARE), which adds anatomy-centered objectives to improve structural preservation in reconstructed volumes. CARE uses a latent diffusion prior pre-trained on 6,212 contrast-enhanced CT scans from 2,870 patients. For a given sparse-view reconstruction method (fixed number of projections), CARE incorporates predictions from multi-structure segmentation models and fine-tunes the diffusion prior using 25 CT scans (20 patients) to learn a reconstruction enhancement module. CARE was evaluated as a post-reconstruction enhancement strategy and compared with four commonly used reconstruction baselines, including a NeRF-based approach.
Results
On the NeRF baseline, anatomy-centered evaluation revealed severe loss of small structures despite acceptable pixel-wise scores, including 3.0% NSD for small organs, 33.2% clDice for vessels, and 1.1% NSD for PDAC. With CARE, structural preservation improved to 53.9% NSD for large organs and 55.5% clDice for vessels, corresponding to gains of +27.9% NSD (large organs), +20.3% NSD (small organs), +22.3% clDice (vascular boundaries), and +3.0% NSD (PDAC) over the original reconstruction.
Conclusion
Pixel-wise reconstruction metrics can overestimate clinical utility under sparse-view acquisition. Anatomy-centered, task-based metrics expose failures in preserving organs, vessels, and tumors. CARE improves structural completeness and provides a practical pathway for radiotherapy-oriented evaluation and enhancement of sparse-view CT reconstruction.