High-Quality Tomographic Image Reconstruction Integrating Neural Networks and Mathematical Optimization

2025-09-07

AI Summary

In nano-/micro-tomography of homogeneous materials, conventional reconstruction methods often produce artifacts and edge blurring at sharp phase boundaries. Method: This paper proposes a neural-enhanced mathematical optimization framework that embeds a lightweight convolutional network—trained to learn local edge priors—into a physics-based iterative reconstruction pipeline, explicitly guiding edge recovery while preserving data fidelity. Unlike end-to-end black-box learning, this hybrid design retains interpretability and generalizability. Results: Experiments on both synthetic and real tomographic data demonstrate significant improvements in boundary sharpness (average +23.6%) and intra-phase uniformity. The method outperforms filtered back-projection (FBP), simultaneous algebraic reconstruction technique (SART), and state-of-the-art deep unrolling approaches in artifact suppression, achieving high-fidelity reconstructions with low blur and strong edge preservation.

Technology Category

Machine Learning: Learning on the Edge & Model CompressionSearch and Optimization: Mixed Discrete/Continuous SearchReasoning under Uncertainty: Stochastic Optimization

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingGraph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphsResponsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

In this work, we develop a novel technique for reconstructing images from projection-based nano- and microtomography. Our contribution focuses on enhancing reconstruction quality, particularly for specimen composed of homogeneous material phases connected by sharp edges. This is accomplished by training a neural network to identify edges within subpictures. The trained network is then integrated into a mathematical optimization model, to reduce artifacts from previous reconstructions. To this end, the optimization approach favors solutions according to the learned predictions, however may also determine alternative solutions if these are strongly supported by the raw data. Hence, our technique successfully incorporates knowledge about the homogeneity and presence of sharp edges in the sample and thereby eliminates blurriness. Our results on experimental datasets show significant enhancements in interface sharpness and material homogeneity compared to benchmark algorithms. Thus, our technique produces high-quality reconstructions, showcasing its potential for advancing tomographic imaging techniques.

Problem

• Enhancing reconstruction quality for homogeneous materials with sharp edges
• Reducing artifacts in tomographic images using neural networks
• Integrating learned edge predictions with mathematical optimization

Innovation

• Neural network trained to identify edges in subpictures
• Integration of trained network into optimization model
• Optimization favors learned predictions with data flexibility