Diffusion Active Learning: Towards Data-Driven Experimental Design in Computed Tomography

Luis Barba, Johannes Kirschner, Tomas Aidukas, Manuel Guizar-Sicairos, Benjamín Béjar

Keywords: Active Learning, Diffusion, Tomography, Computer Vision, Experimental Design

TL;DR: We propose diffusion active learning, an novel framework that integrates generative diffusion models and active learning to reduce data requirements in computed tomography.

Abstract: We introduce _Diffusion Active Learning_, a novel approach that integrates a generative diffusion model with sequential experimental design to adaptively acquire data for solving inverse problems in imaging. We first pre-train an unconditional diffusion model on domain-specific data. The diffusion model is aimed to capture the structure of the underlying data distribution, which is then leveraged in the active learning process. During the active learning loop, we use the forward model of the inverse problem together with the diffusion model to generate conditional data samples from the posterior distribution, all consistent with the current measurements. Based on the generated samples we quantify the uncertainty in the current estimate in order to select the most informative next measurement. We showcase the proposed approach for its application in X-ray computed tomography imaging. Our results demonstrate significant reductions in data acquisition requirements (_i.e._, lower X-ray dose) and improved image reconstruction quality across several real-world tomography datasets.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10594