Research
Representations, priors, and reliable reconstruction.
A closer look at the questions behind my first-author papers.
ECCV2026
We represent a volume as overlapping junctions of 3D wedges, fitting local geometry while encouraging consistency across patches. This training-free prior preserves sharp boundaries in noisy low-dose CT, cryo-electron tomography, and point clouds, and can be used within iterative reconstruction algorithms.
NeurIPS2025
We compare neural, hybrid, and grid representations across 2D and 3D signals at matched parameter budgets. Regularized interpolated grids often fit dense signals faster and more accurately, while neural representations can be advantageous for lower-dimensional structure such as shape contours. The benchmark connects these differences to signal bandwidth, model capacity, and inverse tasks.
IEEE CAMSAP · Invited paper2025
A learned prior can produce a convincing reconstruction even when the target differs from its training data. We use variation across reconstructions from randomized measurements as an instance-level indicator of distribution shift. Experiments on tomographic reconstruction of MNIST digits show increased instability for out-of-distribution targets, without calibration data or retraining.
Preprint2026
We adapt a variational-inequality reconstruction method to polychromatic photon-counting perfusion CT, estimating iodine concentration with a known static background. Digital-phantom experiments investigate how to distribute a limited photon budget across projection views and demonstrate improved reconstruction over filtered back-projection in low-dose settings.