NAB: Neural Adaptive Binning for Sparse-View CT reconstruction

TL;DR

The paper addresses sparse-view CT reconstruction by integrating a shape-aware prior into implicit neural representations. It introduces Neural Adaptive Binning (NAB), a differentiable binning scheme that encodes coordinates into rectangular, rotatable bins whose parameters are learned jointly with a neural network mapping to attenuation coefficients; a projection-domain loss guides end-to-end optimization. NAB’s multi-scale steepness and rotation enable recovery of both rectangular and curved geometries, with theoretical links to hard-bin representations and kernel/binning perspectives. Empirical results on industrial CaCO$_3$ and Workpieces datasets show clear gains over random Fourier-based INRs and classical methods, with robustness demonstrated on medical data and reduced training requirements. The work provides a principled way to inject shape priors into self-supervised CT reconstruction, potentially improving efficiency and generalization in real-world imaging tasks.

Abstract

Computed Tomography (CT) plays a vital role in inspecting the internal structures of industrial objects. Furthermore, achieving high-quality CT reconstruction from sparse views is essential for reducing production costs. While classic implicit neural networks have shown promising results for sparse reconstruction, they are unable to leverage shape priors of objects. Motivated by the observation that numerous industrial objects exhibit rectangular structures, we propose a novel \textbf{N}eural \textbf{A}daptive \textbf{B}inning (\textbf{NAB}) method that effectively integrates rectangular priors into the reconstruction process. Specifically, our approach first maps coordinate space into a binned vector space. This mapping relies on an innovative binning mechanism based on differences between shifted hyperbolic tangent functions, with our extension enabling rotations around the input-plane normal vector. The resulting representations are then processed by a neural network to predict CT attenuation coefficients. This design enables end-to-end optimization of the encoding parameters -- including position, size, steepness, and rotation -- via gradient flow from the projection data, thus enhancing reconstruction accuracy. By adjusting the smoothness of the binning function, NAB can generalize to objects with more complex geometries. This research provides a new perspective on integrating shape priors into neural network-based reconstruction. Extensive experiments demonstrate that NAB achieves superior performance on two industrial datasets. It also maintains robust on medical datasets when the binning function is extended to more general expression. The code will be made available.

Figures (9)

• Figure 1: Evolution of adaptive binning during training using our self-supervised method on two different samples (16 views). The reconstructed CT images gradually approach the ground truth.
• Figure 2: Our method's overall framework.
• Figure 3: Differentiable Binning Features $g(c)_i$ and its generation way in our method (subgraph (a), (b), (c)) and Fixed Fourier Feature in classic INR (subgraph (d)). The $\otimes$ represents Hadamard product, which will act on two 256$\times$256 matrices.
• Figure 4: Differentiable Binning Features $\hat{g}(c)_i$'s multi-scale variations with changing steepness $k_i$.
• Figure 5: Reconstruction at 12, 14, 16 views on the CaCO$_3$ dataset. The best numerical result is marked in red, and the second-best result is marked in blue. Zooming in shows that the comparison methods exhibit edge blurring and distortions, while our method maintains sharp boundaries.