מושגי ליבה
Mitigating data consistency discrepancies in cascaded diffusion models for high-quality image generation in sparse-view CT reconstruction.
תקציר
The content introduces a novel Cascaded Diffusion with Discrepancy Mitigation (CDDM) framework for sparse-view CT reconstruction. It addresses the issue of image degradation in sparse-view CT by utilizing diffusion model-based approaches. The CDDM framework minimizes computational costs and ensures data distribution is close to the original manifold through specialized ADMM and discrepancy mitigation techniques. Experimental results demonstrate superior performance compared to existing methods, emphasizing computational efficiency.
1. Introduction:
- Sparse-view CT reduces radiation exposure but leads to image degradation.
- Diffusion models offer a solution but suffer from training-sampling discrepancy.
2. Background:
- Sparse-view CT reconstruction as an ill-posed inverse problem.
- Standard ADMM and specialized ADMM for TV regularization.
3. Method:
- CDDM framework includes latent and pixel diffusion models with DM and specialized ADMM.
4. Experiments and Results:
- Comparison with benchmarks like DDS, FBPConvNet, and standard ADMM on Walnut and AAPM datasets.
5. Parametric Selection:
- Impact of noise strength, inference steps, and λmax on reconstruction quality.
6. Ablation Studies:
- Effect of DM, gradient directions, and specialized vs standard ADMM on image quality.
סטטיסטיקה
"The initial noisy image depends on the low-quality image x0 and the noise strength t0."
"For ill-posed inverse problems, regularization is commonly introduced to approximate a well-posed problem."
ציטוטים
"Noise perturbation on the low-quality image is crucial for optimal performance."
"Specialized ADMM achieves significantly higher PSNR than standard ADMM."
"DM significantly enhances imaging quality by reducing training-sampling discrepancy."