DuFal: Dual-Frequency-Aware Learning for High-Fidelity Extremely Sparse-view CBCT Reconstruction

TL;DR

DuFal tackles the challenge of extremely sparse-view CBCT reconstruction by introducing a dual-domain encoding framework that jointly leverages spatial structure and high-frequency spectral information. The Frequency Encoder employs HiLocFFNO blocks that combine global (gHiF) and local (lHiF) high-frequency processing, while Spectral-Channel Factorization (SCF) reduces parameter count, and Cross-Attention Frequency Fusion (CAFF) fuses frequency-domain and spatial features. The approach yields superior high-frequency detail preservation and sharper anatomical boundaries, as demonstrated by state-of-the-art PSNR/SSIM and ROI-weighted metrics on LUNA16 and ToothFairy under 6–10 views, along with improved downstream segmentation. DuFal offers favorable efficiency and plug-and-play compatibility with existing reconstruction models, making it well-suited for low-dose CBCT and suggesting broader applicability to other medical imaging modalities when spectral reasoning is advantageous.

Abstract

Sparse-view Cone-Beam Computed Tomography reconstruction from limited X-ray projections remains a challenging problem in medical imaging due to the inherent undersampling of fine-grained anatomical details, which correspond to high-frequency components. Conventional CNN-based methods often struggle to recover these fine structures, as they are typically biased toward learning low-frequency information. To address this challenge, this paper presents DuFal (Dual-Frequency-Aware Learning), a novel framework that integrates frequency-domain and spatial-domain processing via a dual-path architecture. The core innovation lies in our High-Local Factorized Fourier Neural Operator, which comprises two complementary branches: a Global High-Frequency Enhanced Fourier Neural Operator that captures global frequency patterns and a Local High-Frequency Enhanced Fourier Neural Operator that processes spatially partitioned patches to preserve spatial locality that might be lost in global frequency analysis. To improve efficiency, we design a Spectral-Channel Factorization scheme that reduces the Fourier Neural Operator parameter count. We also design a Cross-Attention Frequency Fusion module to integrate spatial and frequency features effectively. The fused features are then decoded through a Feature Decoder to produce projection representations, which are subsequently processed through an Intensity Field Decoding pipeline to reconstruct a final Computed Tomography volume. Experimental results on the LUNA16 and ToothFairy datasets demonstrate that DuFal significantly outperforms existing state-of-the-art methods in preserving high-frequency anatomical features, particularly under extremely sparse-view settings.

Figures (8)

• Figure 1: Illustration of DuFal for CT image reconstruction from multiple projection views. Our proposed Frequency-enhanced Dual-encoding (part a) processes a projection $I_k$ from $K$ input projections through two parallel encoders: a Frequency Encoder that analyzes frequency components of $I_k$ to produce frequency features $\{f_1, f_l, \ldots, f_L\}$ and a Spatial Encoder that produces spatial features $\{s_1, s_l, \ldots, s_L\}$. The model then fuses both domains through a CAFF module before decoding through a series of decoder layers at the Feature Decoder with skip connections to produce the feature $E_k$ of the input image $I_k$. By repeating this process for all projections, we obtain the extracted features $\{E_1, \dots, E_K\}$. The Intensity Field Decoding (part b) operates as in Section \ref{['sec:DIF']}: it first performs Multi-view Fusion by projecting 3D query points to extract view-specific features through bilinear interpolation and then aggregates them using a set function to create unified 3D representations. Subsequently, Intensity Reconstruction employs a multi-layer MLP to decode these fused features into continuous intensity values, which are sampled over a uniform 3D grid to generate the final volumetric CT output $\hat{Y}$.
• Figure 2: Overview of the proposed HiLocFFNO block within the Frequency Encoder. Each HiLocFFNO block comprises two major branches: The gHiF is illustrated at the top using solid blue line (---) and arrows (→), and the lHiF is depicted at the bottom using dashed blue line (- - -) and arrows ($\dashrightarrow$). Both branches employ SCF to replace the full complex weight with separate channel-mixing and spectral-weighting kernels. After inverse Fourier transform, the resulting frequency features $z_l^{\text{glob}}$ and $z_l^{\text{local}}$ are fused with a linear projection $\mathbf{W}$, activated by GeLU, and refined by a Galerkin-Attention (GA) galerkin_att layer to yield the next-level representation $f_{l+1}$.
• Figure 3: The CAFF flowchart. This diagram shows the fusion process between spatial features $s_l$ and frequency features $f_l$ in the frequency domain. Both input features are first convolved and transformed into the frequency domain using Fourier transform $\mathcal{F}$, decomposing into real ($\Re_f$, $\Re_s$) and imaginary ($\Im_f$, $\Im_s$) components. Cross-attention (CA) operations are applied separately to real and imaginary parts, where frequency components serve as keys/values and spatial components as queries. The attended frequency components are combined with original spatial frequency components ($\Re_s$, $\Im_s$) through residual addition $\oplus$, then transformed back to spatial domain via inverse Fourier transform $\mathcal{F}^{-1}$ to produce the fused feature $s'_l$. → indicates the frequency processing path, → shows the spatial processing path, and $\dashrightarrow$ represents the residual connections.
• Figure 4: Visualization of 10-view reconstructed chest CT (from top to bottom: axial, coronal, and sagittal slice). This figure highlights the key advantage of our method in preserving fine details. While DIF-Net and DIF-Gaussian successfully recover the general anatomy, they produce overly smoothed images that lose high-frequency information, as evidenced by the blurry details in the zoomed-in views. In contrast, our method yields visibly sharper reconstructions across the axial, coronal, and sagittal planes, achieving a fidelity much closer to the ground-truth image.
• Figure 5: Visualization of 10-view reconstructed dental CT (from top to bottom: axial, coronal, and sagittal slice). While the competing DIF-Gaussian method reconstructs the general bone structure, it oversmooths the image, completely obscuring fine details like the mandibular canals. Our method yields a significantly sharper result that clearly defines the cortical bone and successfully resolves the mandibular canals in the coronal and sagittal views.