In computed tomography (CT), a 3D volumetric image is reconstructed from a collection of 2D X-ray projections taken at various angles around a subject. Mathematically, each pixel in a projection captures a line integral of the object's attenuation coefficients along the path of the X-ray beam. When hundreds of these projections are acquired, the inverse problem is well-determined, and classical analytical methods like Filtered Backprojection (FBP) can accurately map the data back into a 3D grid.
However, minimizing patient radiation exposure requires reducing the number of projection views. When we transition to a "sparse-view" protocol—sampling only a fraction of the standard angles—the inverse problem becomes severely underdetermined, introducing severe structural artifacts.
The Geometry of Streak Artifacts
Why do classical analytical methods fail so catastrophically in the sparse-view regime? FBP operates by taking the measured attenuation from a detector pixel and smearing it uniformly back along the ray path through the reconstruction space. When dense views are available, these backprojected rays intersect constructively at true anatomical boundaries and destructively cancel each other out in empty space.
Dense Views: Ray intersections cancel out noise -> Clean Reconstruction
Sparse Views: Insufficient rays for cancellation -> Global Streak Artifacts
In a sparse-view setting, there are not enough intersecting rays to achieve this destructive cancellation. Instead, the uncancelled energy manifests as sharp, star-like "streak artifacts" that propagate globally across the entire image. As noted by Kyong Hwan Jin in Deep Convolutional Neural Network for Inverse Problems in Imaging, these artifacts are non-local and highly structured, making them exceptionally difficult to eliminate using standard spatial denoising filters.
The Failure of Discrete Optimization
To overcome the limitations of analytical backprojection, classical algebraic reconstruction techniques discretize the volume into a grid of independent voxels, $x$, and model the projection process as a linear system:
$$Ax = y$$
where $A$ is the forward projection matrix and $y$ is the vector of sparse measurements. In an ultra-sparse setting (e.g., fewer than 20 views), the number of knowns in $y$ is vastly smaller than the number of unknown voxels in $x$.
To stabilize this ill-posed system, traditional methods introduce hand-crafted regularizers, such as Total Variation (TV), which penalizes high gradients to enforce piecewise smoothness. While TV can suppress streaks, it introduces a new problem: "over-smoothing" or "blocky" artifacts that erase fine, low-contrast anatomical structures.
To preserve these details, Peng Bao et al. demonstrated in Sparse-View CT Reconstruction via Convolutional Sparse Coding that replacing patch-based dictionaries with Convolutional Sparse Coding (CSC) allows the system to regularize the image globally through learned feature maps rather than isolated patches. Yet, tuning these hand-crafted or shallow dictionary priors remains computationally expensive and highly sensitive to hyperparameter selection.
Continuous Neural Fields as Implicit Regularizers
The key insight behind recent breakthroughs in sparse-view reconstruction is a fundamental shift in representation: moving away from discrete voxel grids entirely. Instead, the 3D volume is parameterized as a continuous function using an Implicit Neural Representation (INR), or a neural field.
A coordinate-based multi-layer perceptron (MLP), $f_\theta$, is trained to map a continuous 3D spatial coordinate $(x, y, z)$ directly to its corresponding X-ray attenuation value $\mu$:
$$f_\theta(x, y, z) = \mu$$
This continuous formulation changes the nature of the underdetermined problem through the spectral bias of neural networks. Deep networks naturally learn low-frequency, globally consistent structures before fitting high-frequency noise or discontinuities. When optimizing a neural field to match sparse projections, this bias acts as a powerful, continuous regularizer. Instead of producing sharp, discontinuous streak artifacts to fit the sparse data, the network naturally interpolates smoothly across the missing angular gaps.
[Spatial Coordinate (x,y,z)] ---> [ MLP (f_θ) ] ---> [ Attenuation (μ) ]
|
(Differentiable Ray Casting)
|
v
[ Simulated 2D Projection ]
|
(Loss vs. Real Sparse View)
To optimize this network without ground-truth 3D data, researchers employ a self-supervised framework driven by a differentiable forward projection model. For any given X-ray view angle, rays are cast through the virtual volume. The system samples coordinates along each ray, queries the neural field $f_\theta$ to obtain their attenuation values, and integrates them using the Beer-Lambert law to synthesize a simulated 2D projection. Because every step of this ray-casting operation is fully differentiable, the error between the synthesized projection and the actual sparse measurement can be backpropagated directly to update the network weights $\theta$.
In NAF: Neural Attenuation Fields for Sparse-View CBCT Reconstruction, Ruyi Zha et al. demonstrated that combining this continuous formulation with multiresolution hash encoding allows for the rapid recovery of sharp anatomical boundaries from highly limited views without requiring any external training datasets. Similarly, Yiqun Lin et al. leveraged this continuous intensity field concept in Learning Deep Intensity Field for Extremely Sparse-View CBCT Reconstruction to achieve high-resolution 3D Cone-Beam CT (CBCT) reconstructions from as few as 6 to 10 projections.
Deep Unrolling and Hybrid Optimization
While pure neural fields optimize a single volume instance from scratch, another paradigm embeds physical projection models directly into deep neural network architectures through a process known as deep unrolling.
Instead of treating the neural network as a black-box post-processing filter—which Yoseob Han et al. showed in Framing U-Net via Deep Convolutional Framelets: Application to Sparse-view CT can lead to the leakage of high-frequency information—deep unrolling translates classical iterative optimization algorithms (like ADMM or Half-Quadratic Splitting) into a fixed number of network layers. Each layer consists of a physics-based data-fidelity step followed by a learned neural network regularizer.
For example, Yu Guo et al. introduced this hybrid strategy in Deep Inertia $L_{p}$ Half-Quadratic Splitting Unrolling Network for Sparse View CT Reconstruction. By unrolling an $L_p$-norm optimization algorithm and integrating inertial momentum steps, their framework combines the mathematical convergence guarantees of classical optimization with the expressive prior-learning capabilities of deep networks. This hybrid approach ensures that the reconstruction strictly honors the physical forward model of X-ray acquisition while leveraging learned data distributions to fill in missing information.
Data-Driven Priors vs. Physical Consistency
When the number of views drops to extreme scarcity (e.g., fewer than 15 projections), the problem becomes so severely underdetermined that optimization-based interpolation struggles. In these ultra-sparse regimes, data-driven generative priors become necessary to hallucinate the missing anatomy.
For instance, Haofu Liao et al. utilized Generative Adversarial Networks (GANs) equipped with specialized focus maps in Adversarial Sparse-View CBCT Artifact Reduction to clean up severe streak artifacts from highly reduced projection sets. More recently, diffusion models have emerged as a powerful tool for generating highly realistic anatomical details. However, relying heavily on generative priors introduces a critical trade-off between visual plausibility and clinical accuracy.
In a systematic study of this trade-off, Matt Y. Cheung et al. evaluated the utility of diffusion priors in When are Diffusion Priors Helpful in Sparse Reconstruction? A Study with Sparse-View CT. They discovered that while diffusion models provide a significant performance boost in ultra-sparse settings by synthesizing realistic structures where data is completely absent, they quickly hit a performance plateau.
As more projection data becomes available, classical mathematical and physics-constrained optimization methods begin to outperform pure generative models. This highlights a fundamental principle in sparse-view reconstruction: generative priors are invaluable for filling massive informational voids, but physical consistency constraints must take precedence as soon as sufficient measurement data is available to guide the reconstruction.