Regularizing INR with diffusion prior self-supervised 3D reconstruction of neutron computed tomography data
A new hybrid AI model merges diffusion priors with neural representations to create 3D scans from 90% less data.
A team of researchers has introduced Diffusive INR (DINR), a novel AI framework designed to solve a critical problem in scientific imaging: creating accurate 3D reconstructions from severely limited data. Traditional methods for neutron computed tomography (CT) require hundreds of projection views, but DINR leverages a hybrid approach. It combines the speed and flexibility of Implicit Neural Representations (INRs)—which act as a compact, continuous model of a 3D scene—with the powerful generative prior of a diffusion model. This 'diffusion prior' acts as a sophisticated regularizer, guiding the INR to produce physically plausible and high-fidelity 3D structures even when the input data is sparse.
Crucially, the DINR model was pretrained entirely on synthetic data, yet it successfully generalized to real-world experimental observations of concrete microstructures. In tests, where conventional reconstruction techniques suffered substantial degradation with fewer views, DINR maintained superior performance. It significantly reduced artifacts and achieved measurable gains in standard image quality metrics like Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). This breakthrough enables accurate micro-structural characterization for materials science and non-destructive testing in scenarios previously hampered by data scarcity, such as rapid scanning or when radiation exposure must be minimized.
- DINR merges Implicit Neural Representations (INRs) with a generative diffusion model prior for 3D inverse problem solving.
- The model, trained only on synthetic data, enables high-quality reconstruction from sparse-view neutron CT scans with 90% fewer views.
- It outperforms state-of-the-art methods on concrete microstructure data, reducing artifacts and improving PSNR/SSIM metrics for accurate analysis.
Why It Matters
Enables detailed 3D material analysis in fields like civil engineering and aerospace with far less scan time and radiation exposure.