Sinogram upsampling using Primal-Dual UNet for undersampled CT and radial MRI reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Computed tomography (CT) uses ionizing radiation, while magnetic resonance imaging (MRI) has slow acquisition speeds.
• Undersampling in CT and MRI can mitigate these issues but often results in lower resolution and artifacts.
• Existing reconstruction methods typically address CT and MRI separately.

Purpose of the Study:

• To develop a unified deep learning solution for reconstructing sparsely sampled CT and undersampled radial MRI data.
• To improve upon existing deep learning methods for image reconstruction in terms of accuracy and speed.
• To evaluate the proposed method's performance on both CT and MRI datasets, including specific regions of interest.

Main Methods:

• A unified approach using Fourier transform-based pre-processing for radial MRI and sinogram upsampling with filtered back-projection for both modalities.
• Implementation of the Primal-Dual UNet, an enhanced deep learning model based on the Primal-Dual network.
• Validation using fan-beam CT data with a sparsity level of 16 and undersampled brain and abdominal MRI data with an acceleration factor of 16.

Main Results:

• The Primal-Dual UNet achieved a statistically significant improvement in Structural Similarity Index Measure (SSIM) for sparse CT reconstruction (0.932±0.021 vs. 0.919±0.016).
• For undersampled MRI, the model yielded improved average SSIM scores for brain (0.903±0.019 vs. 0.867±0.025) and abdominal data (0.957±0.023 vs. 0.949±0.025).
• The network demonstrated enhanced image quality in regions of interest (liver, kidneys, spleen) and better generalization in the presence of artifacts like needles.

Conclusions:

• The proposed Primal-Dual UNet offers a unified and effective solution for reconstructing undersampled CT and MRI data.
• The method significantly improves image quality and reconstruction speed compared to previous models.
• This unified approach holds promise for advancing non-invasive diagnostic imaging by addressing limitations of current modalities.