Few-View Reconstruction Advances

• Few-view reconstruction is a technique that recovers high-fidelity images and 3D structures from severely sparse data by addressing an underdetermined inverse problem.
• Advanced methods such as total variation, non-convex regularization, and hybrid iterative approaches mitigate artifacts and improve metrics like PSNR and SSIM.
• Deep learning integration enhances reconstruction details by combining physics-based modeling with learned priors, ensuring data consistency in low-dose or rapid-scan scenarios.

Few-view reconstruction denotes the task of recovering high-fidelity images or 3D structures from tomographic or photographic measurements acquired at a severely reduced number of viewing angles or camera positions—often orders of magnitude fewer than classical protocols. This regime is fundamentally ill-posed due to the underdetermined nature of the inverse problem: the available data provide insufficient constraints for unique recovery using conventional analytic methods. Few-view reconstruction is central in low-dose imaging (e.g., CT or SPECT), rapid or fixed-gantry scanning, and modern 3D computer vision where acquisition budgets, motion, or accessibility preclude dense multi-view sampling. Advances in mathematical optimization, regularized inversion, and, more recently, deep learning have enabled substantial progress in mitigating the severe artifacts that characterize naïve reconstructions from sparse data.

1. Mathematical Foundations and Challenges

Let $b \in \mathbb{R}^m$ be the measured data (projections, views, or sinograms), $x \in \mathbb{R}^n$ the image or volumetric density, and $A \in \mathbb{R}^{m \times n}$ the forward (system) operator. In classical settings with $m \gg n$, $A$ is full-rank or overdetermined and analytic inversion methods such as filtered backprojection (FBP) are effective. In the few-view regime ($m \ll n$), $A$ has a large null space, and the system

$b = A x + \varepsilon$

admits infinitely many solutions. Even minimal measurement noise $\varepsilon$ can be dramatically amplified, and reconstructions via conventional methods manifest severe streak artifacts, loss of fine detail, and instability (Kim et al., 2019).

Contemporary approaches address the ill-posedness by introducing prior information or explicitly regularizing the inversion. The general framework is to seek

$$\hat{x} = \arg\min_x \; \Phi_{\mathrm{data}}(x) + R(x),$$

where $x \in \mathbb{R}^n$0 enforces fidelity to the measurements and $x \in \mathbb{R}^n$1 penalizes solutions according to prior constraints—such as sparsity, piecewise smoothness, or learned structure (Champley et al., 2024).

2. Regularization Strategies in Few-View Inverse Problems

2.1 Convex Regularization: Total Variation and Related Penalties

Total Variation (TV) regularization promotes piecewise-constant solutions and is defined as

$x \in \mathbb{R}^n$2

where $x \in \mathbb{R}^n$3 are finite-difference discretizations. TV-based formulations

$x \in \mathbb{R}^n$4

yield robust artifact suppression and have formed the backbone of computational CT for over a decade (Kudo et al., 2016, Champley et al., 2024). Preconditioned primal–dual algorithms and accelerated variants converge rapidly to the TV-minimizer even in the few-view regime, often requiring only a handful of forward/backprojection sweeps (Kudo et al., 2016).

2.2 Non-Convex and Structured Priors

Standard TV can induce over-smoothing, especially in images with fine textures or subtle structures. Innovations include:

• Group-Sparsity Regularization: Captures nonlocal self-similarity by promoting sparsity in groups of similar patches. The GSR-SART algorithm forms groups from overlapping patches and learns adaptive dictionaries per group, leading to superior edge/detail preservation compared to TV (Bao et al., 2018).
• L₀-Norm Gradient Optimization: Enforces sparsity directly on the gradient magnitude image, using iterative hard thresholding (IHT) to preserve only the most significant gradients, outperforming L₁ (TV) in retaining low-contrast or small features (Sun et al., 2014).
• Histogram and Azimuthal Sparsity: For settings in which the expected number of material classes or geometric symmetries are known, sparsity in intensity histograms or azimuthal structure is enforced via non-convex penalties (Champley et al., 2024).

3. Model-Based Iterative and Hybrid Methods

3.1 Multi-Stage and Split Optimization

Optimal few-view reconstruction uses multi-stage pipelines: a low-fidelity seed is computed (e.g., via FBP), followed by a heavy-constraint regularized inversion (e.g., TV, group sparsity), then a final refinement with lighter penalties to recover details without reintroducing null-space artifacts (Champley et al., 2024). Split Bregman and primal–dual schemes efficiently solve subproblems involving non-differentiable regularizers (Bao et al., 2018, Kudo et al., 2016).

3.2 Acceleration and Hybrid Integration

The RISING framework combines an early-stopped rapid iterative solver (for model-based regularization) with a deep learning post-processor. The neural component is trained to learn the "tail" of the iterative sequence, closing the gap between partial convergence and the variational optimum—a paradigm that preserves data-model links and delivers near-optimal results with drastically reduced computation (Evangelista et al., 2022).

4. Deep Learning and Data-Driven Priors

4.1 Deep Priors for Extreme Few-View Regimes

Deep convolutional and transformer-based architectures have been deployed to estimate images directly from few-view sinograms or to act as priors within iterative frameworks. Attention-based CNNs can infer high-frequency structure from sparse data; these inferences then regularize optimization-enforced data-consistency steps (Kim et al., 2019).

4.2 3D Contextual and Volumetric Learning

Fully 3D encoder-decoder networks (e.g., DEAR-3D) correct artifacts across entire volumetric blocks, capitalizing on spatial continuity and the manifestation of streak artifacts in the 3D domain. Adversarial losses (WGAN-GP) further improve realism and texture (Xie et al., 2019).

4.3 Direct End-to-End Learnable Reconstructions

Hybrid architectures, such as DEER, transplant the physics of FBP into neural networks with learnable backprojection modules followed by U-Nets for nonlinearity and detail refinement. These can achieve state-of-the-art detail recovery with orders-of-magnitude fewer parameters compared to conventional deep models (Xie et al., 2019).

5. Quantitative Performance and Comparative Benchmarks

Recent studies report both quantitative and qualitative gains:

Method	PSNR (dB) (V=6)	SSIM (V=6)	PSNR (dB) (V=12)	SSIM (V=12)
FBP	18.5	0.42	21.3	0.55
Regularized Least Sq.	22.8	0.62	25.1	0.72
Sinogram Completion	23.5	0.67	25.8	0.75
Proposed (Deep+Iter.)	25.7	0.74	27.9	0.80

The above table demonstrates that integrating deep inference with regularized iterative reconstruction yields pronounced improvements over both analytic and traditional regularized methods (Kim et al., 2019).

Qualitatively, such approaches produce reconstructions with sharper boundaries, fewer streaks, and reduced noise in homogeneous regions. Quantitative gains in PSNR and SSIM (1–5 dB and 0.07–0.20, respectively) over baselines are consistently observed across varied datasets (Kim et al., 2019, Xie et al., 2019).

6. Limitations and Practical Considerations

• Training and Domain Transfer: The efficacy of learned priors (CNNs, GANs) is contingent on the representativeness of the training set. Out-of-distribution anatomies or acquisition protocols may degrade reconstruction quality (Kim et al., 2019).
• Parameter Selection: Regularization strength (e.g., TV weight $x \in \mathbb{R}^n$5) requires careful tuning. Over-reliance on a learned prior may produce bias; a weak prior can fail to suppress artifacts.
• Computational Cost: Deep or 3D models entail increased memory and compute demand (especially for large volumes) (Xie et al., 2019).
• Data Consistency: Hybrid methods that integrate neural priors within physics-driven optimization maintain data fidelity, but fully end-to-end nets require explicit constraints or strong supervision to prevent "hallucination" of features not supported by the data (Kim et al., 2019).

7. Applications and Outlook

Few-view reconstruction underpins next-generation low-dose, high-speed CT, fixed-gantry architectures (using stationary sources), cardiac and breast imaging, and fast industrial NDT. Model-based and deep-learning paradigms have driven recent improvements from 6–12 view feasibility toward clinical viability without quality compromises (Kim et al., 2019, Xie et al., 2019). Ongoing research explores even more expressive priors, domain adaptation, unsupervised or self-supervised schemes, and integration with uncertainty quantification and diagnostic support.

In summary, few-view reconstruction is a central, active area in tomographic imaging and vision, characterized by a fusion of advanced regularization, optimization strategies, and learned structure priors (Champley et al., 2024, Kim et al., 2019, Xie et al., 2019). Advances in this domain continue to push the envelope on quantitative accuracy and diagnostic utility under extreme data sparsity.