Patient-derived PixelPrint phantoms for evaluating clinical imaging performance of a deep learning CT reconstruction...

Area of Science:

• Medical Imaging
• Radiology
• Computational Imaging

Background:

• Deep learning reconstruction (DLR) algorithms offer advanced image processing capabilities in CT.
• Traditional CT phantoms do not adequately represent complex anatomical structures for evaluating DLR performance.
• Assessing DLR requires realistic phantoms that mimic patient-specific pathologies and anatomy.

Approach:

• A patient-derived 3D-printed PixelPrint lung phantom with ground glass opacities was utilized.
• The phantom was scanned at various radiation doses (0.5-20 mGy) using a conventional CT scanner.
• Images were reconstructed with filtered back projection (FBP), iterative reconstruction, and DLR at multiple denoising levels.

Key Points:

• DLR outperformed FBP and iterative reconstruction across all image quality metrics (noise, CNR, RMSE, SSIM).
• Higher denoising levels in DLR further enhanced performance.
• DLR achieved dose reductions of 25-83% (small phantom) and 50-83% (medium phantom) without compromising image quality.

Conclusions:

• DLR enables diagnostic image quality at significantly reduced radiation doses, enhancing the clinical value of low-dose CT.
• The PixelPrint phantom provides a more realistic evaluation environment for DLR, moving beyond simple noise and contrast assessments.
• This study validates DLR's potential for improving patient safety and diagnostic accuracy in CT imaging.