IntraTomo: Self-supervised Learning-based Tomography via Sinogram Synthesis and Prediction

Guangming Zang, Ramzi Idoughi, Rui Li, Peter Wonka, Wolfgang Heidrich

Abstract

IntraTomo is a self-supervised framework designed for highly ill-posed Computed Tomography (CT) reconstruction problems, such as sparse-view and limited-angle scanning, without requiring external training datasets. The approach integrates a coordinate-based neural network for continuous density field representation with a classical geometry refinement module. By leveraging Fourier Feature Projections (FFP) to mitigate spectral bias and employing a differentiable projection layer, IntraTomo synthesizes missing projection data directly from the available scan. Evaluations on simulated and real-world data demonstrate that IntraTomo significantly outperforms existing methods, achieving superior PSNR and structural fidelity in challenging reconstruction scenarios.

Introduction

Computed Tomography (CT) typically requires dense X-ray projections from a full 360-degree rotation to produce high-fidelity 3D images. However, practical constraints—such as reducing radiation dose or physical obstructions—often lead to sparse-view or limited-angle data. These scenarios result in "ill-posed inverse problems" where traditional mathematical methods like Filtered Back-Projection (FBP) produce blurry, artifact-heavy results. While modern deep learning approaches can mitigate these issues, they typically require thousands of labeled ground-truth scans for training. IntraTomo bridges this gap by learning directly from the single, degraded scan at hand through a self-supervised process.

Methodology

IntraTomo utilizes a hybrid approach consisting of two iterative modules that work in tandem to reconstruct the volume and synthesize missing data.

Sinogram Prediction Module

This module employs a Coordinate-Based Representation (CBR), treating the 3D density field as a continuous function where the input is a $(x, y, z)$ coordinate and the output is the density at that point. To overcome "spectral bias"—the tendency of neural networks to prioritize low-frequency information—the authors implement Fourier Feature Projections (FFP) to map coordinates into a higher-dimensional space. Furthermore, a Differentiable Projection Layer allows the network to re-project its estimated volume back into 2D space, enabling optimization via an L2 loss against the original incomplete input data.

Overview of the IntraTomo framework, showing the sinogram prediction module and the iterative optimization loop.

Geometry Refinement Module

To polish the output of the neural network, this module applies classical optimization techniques. It utilizes Total Variation (TV) to preserve local edges and Non-Local Means (NLM) to maintain global structural consistency. The system iterates between these modules: the network synthesizes missing views to make the problem "well-posed," while the refinement module cleans up spatial details.

Experimental Results

The authors evaluated IntraTomo on a variety of datasets, including simulated phantoms (Shepp-Logan, BrainWeb, ATLAS) and real-world scans (Covid-19 chest CT, botanical specimens).

Key findings include:

  • Superior Fidelity: In limited-angle scenarios (45-degree range), IntraTomo achieved a PSNR of 28.40 dB, significantly higher than the 18.12 dB achieved by Deep Shape Image Prior.
  • Artifact Reduction: In sparse-view tests with only 20 projections, IntraTomo maintained high structural integrity where traditional methods produced severe streaking artifacts.
  • Ablation Insights: The inclusion of the FFP layer alone resulted in a 6.38 dB PSNR improvement, proving essential for capturing high-frequency details. Additionally, the Swish activation function was found to outperform ReLU due to its smoother gradient profile.
Qualitative comparison of sparse-view CT reconstruction across various datasets. IntraTomo (Ours) significantly reduces streaking artifacts compared to traditional and deep learning baselines.

Conclusion

IntraTomo demonstrates that self-supervision and continuous coordinate-based representations are highly effective for medical imaging tasks where external data is scarce. By combining neural synthesis with geometric priors, it achieves state-of-the-art results in sparse and limited-angle tomography. Future work may extend this framework to dynamic tomography for imaging moving objects and applications in Cryo-Electron Microscopy.

Summary updated

Abstract

We propose IntraTomo, a powerful framework that combines the benefits of learning-based and model-based approaches for solving highly ill-posed inverse problems in the Computed Tomography (CT) context. IntraTomo is composed of two core modules: a novel sinogram prediction module, and a geometry refinement module, which are applied iteratively. In the first module, the unknown density field is represented as a continuous and differentiable function, parameterized by a deep neural network. This network is learned, in a self-supervised fashion, from the incomplete or/and degraded input sinogram. After getting estimated through the sinogram prediction module, the density field is consistently refined in the second module using local and non-local geometrical priors. With these two core modules, we show that IntraTomo significantly outperforms existing approaches on several ill-posed inverse problems, such as limited angle tomography with a range of 45 degrees, sparse view tomographic reconstruction with as few as eight views, or super-resolution tomography with eight times increased resolution. The experiments on simulated and real data show that our approach can achieve results of unprecedented quality.
BibTeX
@inproceedings{zang2021intratomo,
  author = {Guangming Zang and Ramzi Idoughi and Rui Li and Peter Wonka and Wolfgang Heidrich},
  title = {{IntraTomo: Self-supervised Learning-based Tomography via Sinogram Synthesis and Prediction}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2021},
  doi = {10.1109/iccv48922.2021.00197},
}
Figure