Computed laminography (CL) is widely used for three-dimensional imaging of plate-shaped objects. However, due to the inherent incompleteness of projection data, CL reconstructions often suffer from severe interlayer aliasing artifacts. In addition, projection truncation often occurs in practical CL imaging scenarios, further degrading reconstruction quality. To address these challenges, we propose a deep learning based framework that integrates a differentiated backprojection (DBP) reconstruction and a Geometry-Robust Artifact Suppression Network (GRASN). The proposed GRASN method explicitly incorporates the geometric characteristics of CL aliasing artifacts into a convolutional neural network (CNN), enabling effective learning of artifact-related features and improved its adaptability to various imaging geometries. Meanwhile, the DBP reconstruction robustly handles truncated projections. Extensive experiments on both simulated and real data demonstrate that the proposed DBP+GRSN method significantly suppresses aliasing artifacts and achieves robust performance across different CL imaging geometries.
Time reversal (TR) serves as a fundamental reconstruction approach for photoacoustic computed tomography (PACT), featuring superior reconstruction accuracy rooted in the rigorous physical process. However, the full-wavefield propagation over the entire computational domain results in low computational throughput. Here, we propose spatial sparsification in the TR process and develop a self-supervised multi-frame super-resolution framework to reconstruct the sparse outputs, thereby achieving video-rate TR imaging without sacrificing image quality. The framework actively degrades images to provide training constraints and then extracts spatiotemporal correlations using 3D convolutions and cascaded structures. Spatiotemporal attention is employed to perform adaptive weighted fusion of multi-frame features. Animal and human finger joint imaging tests verify that our method yields superior image clarity under sparse reconstruction. Compared with conventional TR reconstruction, our method reduces computation time by 93% and achieves a frame rate of ∼24 Hz. When the raw images are used as input, the framework can break through the imaging resolution limit of the hardware and increase the Gray Level Variance by up to 7-fold. Moreover, the accurate image recovery capability of the proposed method enables high-fidelity quantitative analysis of physiological motion. This work overcomes the long-standing throughput bottleneck in TR imaging and is expected to accelerate the clinical translation of high-quality, video-rate PACT imaging.
Magnetic particle imaging (MPI) is an emerging molecular imaging modality that often relies on multipatch scanning to increase the field of view (FOV). Conventional multipatch reconstruction typically requires system matrices (SMs) measured not only inside the FOV but also in surrounding overscan regions to suppress boundary artifacts, which leads to an additional calibration burden, especially in three-dimensional imaging. In this work, we propose a structure-constrained deep learning framework for SM extrapolation using measurements acquired only within the FOV. Three MPI-motivated structural regularizers, namely, local smoothness, gradient-domain regularity, and reflection-equivariance consistency, are incorporated into training. In particular, in the reflection-equivariance term, a dual-input mechanism is used to reduce the spurious sensitivity of the extrapolation mapping to reflection-induced changes in the input orientation, thereby promoting stable and structurally plausible SM extrapolation. Furthermore, a hybrid Uformer architecture with coordinate-aware spatial position embedding and transformer-based skip fusion is used to model both local spatial structure and long-range feature interactions. The results of experiments on simulation, OpenMPI, phantom, and in vivo datasets reveal that the proposed method effectively suppresses boundary artifacts and improves reconstruction fidelity over the evaluated baselines. By avoiding overscan calibration, the proposed framework reduces the number of calibrated SM samples required by 55.6%–81.7% in the studied settings while maintaining competitive reconstruction quality, which suggests a promising calibration-efficient strategy for large-FOV MPI.
Propagation-based phase-contrast CT (PBI-CT) offers significantly enhanced soft-tissue contrast through phase retrieval. Conventional phase retrieval methods typically rely on the homogeneity assumption—that the ratio of the refractive-index decrement to the absorption index ($\kappa = \delta /\beta$) remains constant—which is fundamentally violated in complex imaging objects containing both soft tissue and high-density structures such as bone. In this work, we present a two-step iterative method to decouple the phase contributions of distinct materials. By deriving a material path length estimation theory, we enable the direct calculation of one material's projected path length given the measured propagation intensity and a preliminary estimate of the other material. Our approach combines initial phase retrieval and soft segmentation of the high-attenuation structure with recovery of the remaining material path length from the original projections. Validation through numerical simulations and physical experiments on a laboratory PBI-CT system demonstrates reduced over-smoothing and boundary artifacts relative to conventional single-material phase retrieval. Quantitative analysis shows improved reconstruction fidelity, reducing the root-mean-square error (RMSE) by approximately 87% and increasing the structural similarity index (SSIM) by up to 20% relative to conventional Paganin retrieval. In addition, numerical simulations confirm that the proposed method retains the advantage of higher contrast-to-noise ratio (CNR) of phase contrast CT for low-contrast soft-tissue structures compared with absorption CT. The resulting framework provides a physics-based approach to multi-material artifact reduction in X-ray phase-contrast tomography.
Single-photon LiDAR enables high-quality depth and intensity reconstruction from low photon budgets, but recovery in challenging environments is constrained by extreme signal-to-background ratios (SBR), spatially heterogeneous noise, and the Poisson statistics of photon detection. Existing methods typically operate in Euclidean space and assume spatially uniform noise; consequently, they do not account for the Fisher–Rao geometry of Poisson observations and have limited adaptability to spatially varying backgrounds. We introduce Fisher–Rao Adaptive Reconstruction (FAR), a strongly convex optimization framework that combines geometrically consistent reconstruction with SBR-adaptive anisotropic regularization. A square-root reparameterization flattens the Poisson Fisher–Rao metric to a constant multiple of the Euclidean metric and makes the data-fidelity term uniformly strongly convex without explicitly computing the Fisher information matrix. A frozen-weight mechanism computes spatial conductances from the initial estimate, enabling data-driven adaptation while maintaining convexity. Experiments on Middlebury scenes demonstrate that FAR maintains consistent reconstruction performance under extreme SBR and across spatial heterogeneity patterns, preserving fine structures and sharp edges where competing methods oversmooth or introduce artifacts. Key metrics such as depth RMSE, intensity PSNR, and the no-reference image-quality metric NIQE confirm robust reconstruction across these conditions. Real-world experiments under steady and flickering illumination show reduced streak artifacts and improved detail recovery in the tested practical conditions.
