Ultra-Low-Dose CTPA Using Sparse Sampling CT Combined with the U-Net for Deep Learning-Based Artifact Reduction: An...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Computed Tomography

Background:

• Dose reduction in CT pulmonary angiography (CTPA) is crucial but can compromise image quality due to artifacts.
• Sparse-sampling CT (SpSCT) offers dose reduction but introduces significant artifacts.
• Deep learning, specifically U-Net architectures, shows promise for image reconstruction and artifact reduction.

Purpose of the Study:

• To evaluate the efficacy of a U-Net-based approach for artifact reduction in dose-reduced SpSCT.
• To assess the impact of U-Net post-processing on image quality and diagnostic performance in CTPA.
• To compare U-Net performance against traditional filtered back projection (FBP) and evaluate automated pulmonary embolism (PE) detection.

Main Methods:

• A dual-frame U-Net model was trained on 69 patient datasets and validated on 16.
• SpSCT data were generated from 89 patients with varying views (16 to 512).
• Image quality was assessed using structural similarity index (SSIM) and a reader study; diagnostic performance was evaluated using Sørensen-Dice coefficient and automated PE detection.

Main Results:

• U-Net significantly improved image quality across all view subsets (e.g., SSIM increased from 0.378 to 0.892 at 64 views).
• Reader studies confirmed enhanced image quality, increased diagnostic confidence, and reduced artifacts with U-Net post-processing (P < 0.05).
• Diagnostic performance improved significantly for 64- and 32-view images (Sørensen-Dice: 0.44 vs. 0.23 and 0.09 vs. 0.00, respectively).

Conclusions:

• U-Net-based artifact reduction effectively enhances image quality and diagnostic performance in low-dose SpSCT.
• The findings support the potential for substantial radiation dose reduction in CT pulmonary angiography.
• U-Net post-processing offers a viable solution for improving the diagnostic utility of sparse-sampling CT.