Investigates how including unlabeled data and meta-ablation studies improves the validation of simulation-based evaluation methods for medical imaging models.

Proposes a Fourier neural operator conditioned on the acquisition trajectory to reconstruct cone-beam CT images from non-circular source orbits.

Compares lightweight model architectures for translating text embedded in document images, evaluating the trade-off between model size and translation quality.

Studies incremental learning strategies for classifying document types from a combination of textual content and layout features as new document classes are introduced over time.

Introduces CircleID, a large-scale ICDAR 2026 competition using a new 46,155-image dataset of hand-drawn circles to benchmark open-set writer identification and cross-writer pen classification across hundreds of participating teams.

Proposes a differentiable, graded relaxation of Tuy's completeness condition, together with an Effective Spatial Resolution diagnostic, for projection selection in region-of-interest cone-beam CT, and shows via NP-completeness analysis that a submodular greedy algorithm with proven approximation guarantees nearly matches an exact MILP solution.

Analyzes the robustness of a differentiable shift-variant FBP framework for cone-beam CT, showing it remains stable under irregular and non-planar trajectories, tolerates sparse-view acquisition with far lower computational cost than iterative methods, and is more sensitive to spatial sampling distribution than to trajectory ordering.

Presents diffct, a CUDA-accelerated, open-source library providing differentiable forward and exact adjoint operators for 2D parallel-beam, 2D fan-beam, and 3D cone-beam CT, extending from a stable circular-orbit release to arbitrary per-view source/detector trajectories.

Proposes a differentiable, end-to-end trainable approximate-truncation-robust (ATRACT) CBCT reconstruction network built via known operator learning that optimizes redundancy weights under limited-angle/truncated geometry, improving SSIM and MSE over Parker-weighted analytical reconstruction.

Presents Filter2Noise-4D, a zero-shot, interpretable content-adaptive bilateral filtering framework with only about 1.8k parameters that exploits spatio-temporal correlations across neighboring slices to denoise 4D low-dose CT without paired training data or per-protocol retraining.

Presents Filter2Noise, a self-supervised, interpretable low-dose CT denoising framework built on an attention-guided bilateral filter with a multi-scale loss and Euclidean Local Shuffle training scheme, achieving state-of-the-art zero-shot performance on Mayo Clinic and photon-counting CT data with only about 3.6k parameters.

Proposes a test-time, physics-guided manifold optimization approach to blindly compensate for motion artifacts in single-scan cone-beam CT without requiring external motion sensors or paired training data.

Introduces a motion-aware joint bilateral filter guided by deformable attention to denoise 4D CT data in a zero-shot setting while accounting for inter-phase motion.

Proposes Noise2Flow, a single-image denoising method that learns a conditional mean denoising flow from weak anchor couplings instead of paired clean/noisy training data.

Describes an updated PYRO-NN library for differentiable CT reconstruction, adding PyTorch compatibility, native CUDA kernels for parallel/fan/cone-beam projection and back-projection, artifact-simulation tools, arbitrary trajectory modeling, and a high-level API for end-to-end trainable pipelines.

Introduces a neural network for cone-beam CT reconstruction from non-circular trajectories that replaces the trajectory-dependent filtering component of a differentiable shift-variant FBP model with a trainable 2D Gaussian model, cutting parameters by about 99% and reducing per-trajectory training time to roughly a quarter of the original while preserving reconstruction quality.

Introduces an enhanced, interpretable FDK-based neural network for 3D cone-beam CT that uses wavelet transforms to sparsify trainable cosine-weighting and filtering parameters, cutting parameter count by 93.75% while preserving reconstruction quality and the classical FDK's inference cost.

Proposes standardizing non-circular cone-beam CT source trajectories so that a single differentiable shift-variant FBP model, trained once, generalizes across trajectory types instead of requiring per-trajectory retraining.

Introduces LSTT, an end-to-end framework that uses a VQ-VAE to tokenize CBCT projections, a temporal transformer to model motion dynamics, and a differentiable FDK layer to directly correct non-rigid respiratory motion from projection data without external gating hardware.

Presents a model-agnostic framework that applies 2D generative models slice-by-slice with a through-plane total-variation consistency loss to compensate for motion artifacts caused by mechanical vibration in industrial cone-beam CT.

Proposes a two-step few-view CT reconstruction pipeline that first uses the Discrete Algebraic Reconstruction Technique to generate a binary structural prior, then integrates that prior into a differentiable known-operator-learning framework via gradient-update cropping and prior-informed initialization to reconstruct from sparse, arbitrary-trajectory projection data.

Proposes two deep-learning-enhanced alternatives to iterative reconstruction for arbitrary CBCT trajectories — one optimizing shift-variant filtering before reconstruction and one applying a learned filter afterward — that improve reconstruction quality over iterative methods while reducing computation.

Proposes optimizing CT scan trajectories using a ResNet-18 model trained to predict defect-detection probability from projections of CAD volumes with embedded artificial defects, then greedily selecting projections that balance high Tuy completeness against task-based defect visibility.

Applies PCA to the redundancy weights learned by a differentiable shift-variant FBP model for non-circular CBCT trajectories, achieving a 97.25% reduction in trainable parameters without loss of reconstruction accuracy and substantially faster training.

Shows that a differentiable shift-variant FBP neural network can learn redundancy weights for non-continuous or randomly ordered cone-beam CT trajectories, indicating that data-driven weight learning depends on spatial sampling rather than strict trajectory ordering.

DRACO introduces a differentiable reconstruction framework for cone-beam CT (CBCT) that supports arbitrary imaging orbits, enhancing flexibility and accuracy in CT reconstruction.

EAGLE proposes a novel edge-aware loss function designed to enhance gradient localization, improving the preservation of edge details in CT image reconstruction.

This paper explores various approaches to modeling measuring systems, ranging from traditional white-box models to advanced cognitive methodologies, highlighting their applications and limitations.

Proposes a reinforcement learning framework combining a modified artificial potential field method with Deep Deterministic Policy Gradient and a tailored reward with a compensation term, achieving safer, more energy-efficient obstacle avoidance than baselines including a TD3 variant.

This paper introduces a gradient-based approach for fast and accurate compensation of head motion in cone-beam CT, improving reconstruction quality under motion scenarios.

An overview of data-driven approaches in metrology, highlighting recent advancements and potential future directions.

Explores AI-driven automation techniques for robotic CT scanning procedures to improve efficiency and reliability.

This paper introduces an integer optimization approach for CT trajectories, leveraging a discrete data completeness formulation to enhance imaging performance and efficiency.

Presents a novel sinogram-based technique for localizing defects in CT images, improving diagnostic accuracy.

Investigates the use of gated recurrent units (GRUs) to optimize CT trajectories for enhanced image quality and reduced scanning times.

Introduces trainable Fourier series for filter design in filtered back-projection CT reconstruction, achieving improved results over conventional methods.

Explores deep learning applications to CT reconstruction, building on the Defrise and Clack algorithm for improved artifact suppression.

Presents a gradient-based learning approach for optimizing fan-beam CT geometries, enhancing image quality and reducing artifacts.

Describes a task-based approach for generating optimized CT projection sets using differentiable ranking methods.

This paper discusses methods for optimizing CT scan geometries, comparing gradient-based and non-gradient-based approaches to improve imaging performance and reduce artifacts.

Develops a learning-based optimization framework for twin robotic CT systems, enhancing operational efficiency and imaging results.