Abstract
Synchrotron X-ray tomography enables the examination of the internal structure of materials at submicron spatial resolution and subsecond temporal resolution. Unavoidable experimental constraints can impose dose and time limits on the measurements, introducing noise in the reconstructed images. Convolutional neural networks (CNNs) have emerged as a powerful tool to remove noise from reconstructed images. However, their training typically requires collecting a dataset of paired noisy and high-quality measurements, which is a major obstacle to their use in practice. To circumvent this problem, methods for CNN-based denoising have recently been proposed that require no separate training data beyond the already available noisy reconstructions. Among these, the Noise2Inverse method is specifically designed for tomography and related inverse problems. To date, applications of Noise2Inverse have only taken into account 2D spatial information. In this paper, we expand the application of Noise2Inverse in space, time, and spectrum-like domains. This development enhances applications to static and dynamic micro-tomography as well as X-ray diffraction tomography. Results on real-world datasets establish that Noise2Inverse is capable of accurate denoising and enables a substantial reduction in acquisition time while maintaining image quality.
Introduction
Synchrotron X-ray tomography serves as a high-resolution "microscope" for the internal structure of materials. However, researchers often face a fundamental trade-off: capturing fast-moving processes or protecting sensitive samples requires short exposure times, which inevitably leads to noisy images. While deep learning models like Convolutional Neural Networks (CNNs) are highly effective at denoising, they traditionally rely on supervised training with pairs of noisy and "clean" reference images. In many synchrotron experiments, such clean references are impossible to obtain. This paper presents an expansion of Noise2Inverse, a self-supervised method that denoises multi-dimensional data without any high-quality reference.
The Noise2Inverse Methodology
The core of the Noise2Inverse approach is a clever use of statistical independence. By splitting a single noisy measurement into independent subsets, the network can learn to identify the common underlying structure while ignoring uncorrelated noise.
As illustrated above, the process involves:
- Sinogram Splitting: Partitioning the raw data into independent subsets.
- Independent Reconstruction: Creating multiple noisy versions of the same object.
- Self-Supervised Training: Training a CNN to map one noisy version to another. Because the noise is random but the object is constant, the network learns to reconstruct only the signal.
Multi-dimensional Extensions
This research extends the 2D Noise2Inverse concept into higher dimensions to better capture the complexities of synchrotron data:
- Space (2.5D): Uses adjacent slices as context to improve 3D structural consistency.
- Time (Dynamic): Employs interlaced sampling to denoise sequences of moving objects.
- Spectrum (XRD-CT): Treats different diffraction angles as channels to learn chemical phase interdependencies.
The authors utilized a Mixed-Scale Dense (MS-D) network architecture, which is highly parameter-efficient, allowing it to handle large datasets or very small ones (like the 10 MB archaeological ceramic dataset) with the same hyperparameters.
Key Results and Insights
The method was validated on real-world datasets, including fuel cells and archaeological artifacts.
- Efficiency: The approach allows for a 70-80% reduction in acquisition time while maintaining high image quality.
- Quality: Noise2Inverse produced sharper images with fewer artifacts than standard methods like Total-Variation (TV) minimization, avoiding common issues like "stair-casing."
- Generalizability: The same network configuration worked across vastly different domains, suggesting a robust "launch and forget" tool for synchrotron scientists.
Conclusion
By leveraging the inherent structure of multi-dimensional data, Noise2Inverse removes the primary barrier to deep learning in tomography: the need for clean training data. This advancement enables faster, lower-dose imaging across a wide range of scientific disciplines.
Summary updated