Sparse-view synchrotron X-ray tomographic reconstruction with learning-based sinogram synthesis.

Area of Science:

• Medical Imaging
• Computational Biology
• Materials Science

Background:

• Synchrotron X-ray microscopy enables high-resolution imaging for tomography.
• Acquiring numerous projections for detailed tomography is time-consuming, costly, and increases radiation exposure.
• Existing sparse acquisition methods often yield images with artifacts and noise.

Purpose of the Study:

• To develop a deep-learning-based approach for tomographic reconstruction using sparse-view X-ray projections.
• To address challenges of time consumption, high cost, and radiation dose associated with dense X-ray imaging.
• To improve the quality of tomographic reconstructions from limited projection data.

Main Methods:

• A convolutional neural network (CNN) interpolates sparse X-ray projections to create a dense sinogram.
• A second CNN is employed for error correction in the reconstructed sinogram.
• Transfer learning was utilized to adapt a model trained on Drosophila data to improve mouse tomography reconstruction.

Main Results:

• The proposed deep learning method successfully generated high-quality tomography images from sparse-view projections.
• The approach demonstrated effectiveness on both Drosophila and mouse datasets.
• Transfer learning significantly enhanced the reconstruction quality for the smaller mouse dataset.

Conclusions:

• Deep learning, specifically CNNs, offers a viable solution for high-quality sparse-view tomography reconstruction.
• This method can substantially reduce imaging time, cost, and radiation dose in synchrotron-based X-ray microscopy.
• The application of transfer learning improves model performance on smaller or related datasets.