Summary
This paper frames reconstruction as an optimization problem regularized by a deep generative network rather than a supervised training dataset. The network acts as an implicit structural prior, allowing the method to suppress noise and preserve anatomy without relying on paired examples.
Instead of learning a mapping from low-dose images to clean targets across a large dataset, the method optimizes a randomly initialized convolutional generator directly for the reconstruction instance at hand. That places the work closer to deep image prior style reconstruction than to standard data-driven denoising pipelines.
Method
The reconstruction objective combines measurement consistency with image-domain regularization terms, so the solution is pushed to explain the observed projections while remaining structurally plausible. The paper studies different loss variants and compares them against conventional filtered back-projection, SART, and TV-regularized SART baselines on both analytical phantoms and human CT data.
Why it matters
This paper is an early lab result showing that neural networks can contribute to reconstruction even when no training dataset is available. That matters in medical imaging because clean, paired target data is often difficult to obtain, especially when a new scanner setting, dose regime, or protocol is being evaluated.
Highlights
- Uses an untrained or lightly structured network as a reconstruction prior.
- Avoids the requirement for a dedicated clean target dataset.
- Connects classical reconstruction objectives with deep generative regularization.