A new artificial intelligence framework called LUCID (Learned Undersampling-Adaptive Consistency-Guided Inference with Deterministic Flow Matching) addresses the challenge of reconstructing high-quality CT images from sparse-view data, according to a paper published on arXiv. Sparse-view CT acquisition reduces radiation dose and scanning time by capturing fewer projection views, but the resulting angular undersampling makes reconstruction severely ill-posed, leading to streak artifacts, structural blurring, and loss of fine details.
The Problem of Sparse-View CT
Existing supervised deep learning methods for sparse-view CT reconstruction are often tied to specific sampling settings, limiting their generalization. Generative methods, meanwhile, may introduce anatomically inconsistent hallucination-like structures under severe undersampling. The LUCID framework aims to overcome these limitations by providing a sparsity-adaptive, consistency-guided reconstruction approach that works across different sampling densities.
LUCID's Approach: Flow Matching with Adaptive Consistency
LUCID is based on a Flow Matching generative prior. The model is trained only on high-quality CT images to learn a continuous transport between a Gaussian distribution and the high-quality CT image distribution, independent of the view sampling pattern. This training approach makes the prior decoupled from the specific undersampling geometry.
During inference, LUCID explicitly incorporates the sampling sparsity level to adapt the generative trajectory of the single pretrained model. The process involves three key steps:
- Degradation-matched initial state construction: The initial state is created by a sparsity-weighted fusion of the sparse-view FBP (filtered back projection) image and Gaussian noise.
- Sparsity-modulated Flow Matching updates: The generative trajectory is modulated by the sparsity level.
- Projection-domain data-consistency correction: After each prior update, a correction step ensures consistency with the measured projection data.
This approach allows LUCID to adapt to varying undersampling conditions without retraining the model.
Experimental Validation
Experiments conducted under multiple sparse-view settings demonstrated that LUCID achieves stable reconstruction performance across different sampling densities. According to the paper, the method improves image quality and structural fidelity and reduces the risk of hallucination-like structures compared to other generative sparse-view CT reconstruction methods. The authors reported that LUCID effectively handles the ill-posed nature of sparse-view reconstruction while maintaining high fidelity to the underlying anatomy.
The table below summarizes the key components of LUCID versus conventional approaches:
| Feature | Conventional Supervised Methods | LUCID |
|---|---|---|
| Training data requirement | Paired sparse-view and full-view CT images | High-quality CT images only |
| Adaptability to different sampling | Tied to specific sampling settings | Sparsity-adaptive via inference modulation |
| Artifact handling | Limited by training distribution | Consistency correction with flow matching |
| Hallucination risk | Moderate | Reduced via deterministic flow and correction |
Implications for Enterprise Healthcare
For technology leaders in healthcare, LUCID represents a step toward more flexible and clinically deployable AI models for CT reconstruction. By reducing the need for paired training data and enabling adaptation to different scanner protocols, the framework could lower the barrier for integrating advanced reconstruction algorithms into existing imaging workflows. The ability to reduce radiation dose without sacrificing image quality has direct implications for patient safety and operational efficiency in radiology departments. Further research and clinical validation will determine how quickly such methods can be adopted in production environments.
The research was authored by Duan, Jigang, Wang, Jiayi, Heran, Yang, Ping, Ma, Genwei, and Zhao, Xing. The full paper is available on arXiv under the identifier 2606.16212.