Score-based Random Tomography of Dynamic Structures

Julian Leander Möbius, Michael Habeck

TL;DR: Random tomography of flexible 3D structures without known orientations using 2D score-based diffusion.

Abstract: A central problem in imaging sciences is the reconstruction of a three-dimensional object from two-dimensional projections. In many practical settings, however, the orientations at which projections are acquired cannot be controlled or observed. An additional challenge arises when the object is not static but undergoes continuous structural changes during data acquisition, a common characteristic of biomolecular systems imaged by cryo-electron microscopy. We propose a two-stage approach to random tomography of flexible objects that exhibit such structural variability. First, we learn a score-based diffusion model directly from 2D projection images, capturing the distribution of object conformations under unknown and randomly distributed orientations. Second, we reconstruct 3D volumes whose projections are consistent with the probability distribution implicitly defined by the learned diffusion model. Combined, these two stages enable 3D reconstruction in the presence of both unknown orientations and intrinsic structural dynamics.

Primary Area: Applications->Computer Vision

Keywords: random tomography, score-based diffusion model, score distillation sampling, structural heterogeneity, cryo-electron microscopy, 3D reconstruction

Submission Number: 9795