SimAQ: Mitigating Experimental Artifacts in Soft X-Ray Tomography using Simulated Acquisitions

TL;DR

SimAQ addresses the challenge of artifact-laden soft X-ray tomography (SXT) data and scarce annotations by introducing a 3D synthetic data pipeline that generates realistic yeast phantoms, applies controlled imaging artifacts, and produces paired noisy reconstructions with ground-truth segmentations. A hybrid 3D-to-2D U-Net is trained predominantly on synthetic data and tuned with limited real annotations, enabling robust artifact correction and segmentation under limited-angle, high-artifact conditions. The approach demonstrates high segmentation accuracy on synthetic tests and effective zero-shot and few-shot transfer to real tomograms, reducing reliance on large labeled datasets. This work provides a practical framework for systematic benchmarking and development of SXT analysis methods with broad applicability to artifact-prone tomographic modalities.

Abstract

Soft X-ray tomography provides detailed structural insight into whole cells but is hindered by experimental artifacts such as the missing wedge and by limited availability of annotated datasets. We present SimAQ, a simulation pipeline that generates realistic yeast phantoms and applies synthetic imaging artifacts to produce paired noisy volumes, sinograms, and reconstructions. We validate our approach by training a neural network primarily on synthetic data and demonstrate effective few-shot and zero-shot transfer learning on real X-ray tomograms. Our model delivers accurate segmentations, enabling quantitative analysis of noisy tomograms without relying on large labeled datasets.

Figures (8)

• Figure 1: Examples of experimental artifact sources in cryo-SXT. A) Motion blurring resulting in missing projections. B) Flat field mismatch. The yellow flat field correction on top of the acquired projection (black), is often based on flat field images samples from elsewhere. C) Cracks in the ice. D) High absorption fiducial markers. E) Inconsistent ice attenuation (the red dashed line is 2.4 times further than the green in this example). F) Detector dark current. Please note that the objects are not to scale and that the X-ray source and detector are stationary, while the specimen is rotated.
• Figure 2: Projections from a real tomogram of a yeast cell acquired at Bessy II (left) and from a synthetic tomogram with comparable structural features (right). The plotted intensity profile represents the mean pixel value within the red rectangle across all projection angles.
• Figure 3: Hybrid 3D-to-2D U-Net architecture. The 3D encoder extracts volumetric features across multiple scales using slice-wise 2D pooling. The central slice from the bottleneck layer feeds a 2D decoder with skip connections, enabling 3D context-aware segmentation with 2D computational cost.
• Figure 4: Representative examples from the synthetic evaluation dataset. Top row: noisy input reconstructions. Bottom row: corresponding ground truth segmentations with cell membrane in blue, vacuole in red and lipid droplets in green.
• Figure 5: Violin plot of IoU scores for 100 sampled tomograms across different cell counts and channels. Reconstructions are performed with an $100\degree$ projection span. Background class has been omitted since all scores are $>0.99$