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

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Deep learning reconstruction (DLR) algorithms have object-dependent performance, challenging traditional phantom evaluations.
• Existing CT phantoms may not accurately represent clinical imaging scenarios for DLR.

Purpose of the Study:

• To evaluate a commercial DLR algorithm using a patient-derived 3D-printed lung phantom.
• To assess DLR performance across various radiation dose levels and phantom sizes.
• To compare DLR against traditional reconstruction methods (FBP, iterative reconstruction).

Main Methods:

• A 3D-printed lung phantom with realistic pathologies (ground glass opacities) was created using PixelPrint technology.
• Scans were performed on a conventional CT scanner at doses ranging from 0.5 to 20 mGy.
• Images were reconstructed using FBP, iterative reconstruction, and DLR with five denoising levels.

Main Results:

• DLR outperformed FBP and iterative reconstruction across all image quality metrics (noise, CNR, RMSE, SSIM, MS SSIM).
• DLR demonstrated superior performance with increased denoising levels.
• Estimated dose reduction with DLR ranged from 25%-83% without compromising image quality.

Conclusions:

• DLR can achieve diagnostic image quality at significantly reduced radiation doses (up to 83%).
• The PixelPrint phantom provides a more realistic evaluation environment than traditional phantoms.
• DLR enhances the clinical utility of low-dose CT scans.