A convolutional neural network for fast upsampling of undersampled tomograms in X-ray CT time-series using a...

Area of Science:

• * X-ray computed tomography (CT)
• * Image processing and analysis
• * Deep learning applications in scientific imaging

Background:

• * Time-resolved volumetric tomography (4D datasets) generates large amounts of data, often requiring compromises in projection numbers for sufficient temporal resolution.
• * Low projection counts per tomogram can limit image quality and hinder downstream analysis, such as feature detection and segmentation.
• * Existing interpolation techniques struggle to accurately reconstruct high-quality tomograms from limited projection data.

Purpose of the Study:

• * To develop a deep learning-based super-resolution method for upscaling sinogram data in 4D CT datasets.
• * To improve the quality of tomographic reconstructions from sparsely projected data.
• * To enable more accurate feature detection and segmentation in time-resolved imaging studies.

Main Methods:

• * A deep neural network (UDNN) is proposed, trained to learn an end-to-end mapping between low- and high-projection sinograms.
• * The network utilizes prior information from highly sampled tomograms to accurately upscale the sinogram space.
• * A lightweight convolutional neural network architecture is designed for efficient retraining across different sample types.

Main Results:

• * The UDNN approach demonstrates superior accuracy in upscaling sinograms compared to traditional interpolation methods.
• * Reconstructed tomograms show increased quality, particularly for datasets with limited projections, facilitating easier segmentation.
• * The method was validated on both synthetic and real-world experimental data, including released large-volume datasets.

Conclusions:

• * The proposed deep learning method effectively enhances the quality of time-resolved X-ray CT reconstructions.
• * Upscaling sinogram data significantly aids in analyzing dynamic processes captured at high frame rates.
• * This technique offers a valuable tool for researchers requiring high-fidelity 4D imaging and analysis.