Self-supervised denoising of raw tomography detector data for improved image reconstruction

TL;DR

The paper addresses denoising of noisy raw sinogram data from ultrafast X-ray CT to improve image reconstruction. It evaluates self-supervised denoising models Noise2Void (N2V) and Noise2Void2 (N2V2) against BM3D on a calibrated phantom dataset, using five-fold cross-validation and PSNR as the metric. N2V delivers the most consistent improvements, with median reconstruction-space gains around $+4.1$ dB and $+3.3$ dB in sinogram space, while N2V2 shows more variable results and BM3D underperforms. Denoising the raw detector data also yields improved reconstructions, supporting broader validation on diverse datasets and downstream tasks.

Abstract

Ultrafast electron beam X-ray computed tomography produces noisy data due to short measurement times, causing reconstruction artifacts and limiting overall image quality. To counteract these issues, two self-supervised deep learning methods for denoising of raw detector data were investigated and compared against a non-learning based denoising method. We found that the application of the deep-learning-based methods was able to enhance signal-to-noise ratios in the detector data and also led to consistent improvements of the reconstructed images, outperforming the non-learning based method.

Figures (7)

• Figure 1: Overview of our training pipeline as detailed in section \ref{['sec:methods']}. We trained two machine learning based denoising models (N2V and N2V2) using a cross-validation strategy and compared to a non-neural-network-based denoising approach (BM3D).
• Figure 2: PSNR difference ($\text{PSNR}_{denoised} - \text{PSNR}_{original/noisy}$) for N2V, N2V2, and BM3D across five-fold cross-validation in both sinogram and reconstruction domains. N2V shows stable improvements, N2V2 is more variable, and BM3D consistently underperforms.
• Figure 3: Example reconstructed slice from the best-performing fold of N2V2, showing effective noise suppression while preserving structural details.
• Figure 4: Example reconstructed slice from the best-performing fold of N2V, demonstrating consistent denoising and structural preservation.
• Figure 5: Example reconstructed slice from BM3D denoising, illustrating its filtering effect on the noisy input.