We propose the first flow-matching-based framework for sparse-view CT reconstruction. Motivated by the strong temporal correlation of velocity fields across time steps, we introduce an efficient variant that reuses previously predicted velocities to significantly reduce computational cost.

DM4CT is a comprehensive benchmark for diffusion-based methods in CT reconstruction. It systematically compares ten state-of-the-art diffusion models alongside modern learning-based approaches (e.g., Gaussian Splatting and INRs) and classical reconstruction algorithms across diverse CT datasets. The benchmark provides detailed insights into reconstruction quality, efficiency, and practical trade-offs. In addition, we introduce a new real-world high-resolution synchrotron CT dataset to facilitate realistic evaluation of CT reconstruction methods.

A self-supervised approach designed for edge illumination X-ray phase contrast micro-CT to enhance the resolution of anisotropic volumes.

A artifact reduction method for CT images based on deep learning. Three CNNs reduce artifacts in a subsequent manner, targeting different types of artifacts. This method is seamlessly integrated into existing CT pipelines.

A novel Bayesian framework for joint reconstruction of multiple objects from sparse-view CT scans using Implicit Neural Representations (INRs) to improve reconstruction quality. By capturing shared patterns across multiple objects with latent variables, our method enhances the reconstruction of each object, increases robustness to noise, and accelerates the learning process.

A study of how the choice of denoising method affects downstream defect detection in fast-scan industrial CT. BM3D, a supervised UNet, the self-supervised Noise2Inverse and a multi-stage CNN are compared on two real aluminium samples, scored both on image quality (PSNR/SSIM) and on the Dice score of the defects segmented from their outputs. Denoisers that look good by image-quality metrics can erase the small, low-contrast pores that matter, showing that the denoiser must be selected for the analysis task and not for the metric alone.

A Benchmark Low-dose Industrial CT Dataset Tailored for Deep Learning: This work introduces a dataset specifically designed for industrial CT applications, emphasizing deep learning approaches. It features a comprehensive analysis comparing the effectiveness and efficiency of 2D, 2.5D, and 3D training methodologies in the context of denoising and image enhancement.

A novel method to reconstruct a spectral central view and its aligned disparity map from spatio-spectrally coded light fields using multi-task deep learning.

Why do over-parameterized networks generalize? From a minimum description length perspective, we propose the prequential regret, computable from the training trajectory alone, and show that it upper-bounds the description length of the trained parameters, which in turn yields a PAC-MDL generalization bound. Empirically, the regret closely tracks the generalization error on CIFAR-10 and CIFAR-100.

Graph neural network for log anomaly detection.

Unsupervised cross-domain graph level anomaly detection.

Survey paper on diffusion models for tabular data.

Survey paper on reinforcement learning for LLM post-training.