Performance of clinically available deep learning image reconstruction in computed tomography: a phantom study

Area of Science:

• Medical Imaging Physics
• Radiological Technology
• Artificial Intelligence in Healthcare

Background:

• Traditional computed tomography (CT) image reconstruction methods like filtered back projection (FBP) and iterative reconstruction (IR) have limitations in noise reduction and radiation dose efficiency.
• Deep learning image reconstruction (DLIR) is an emerging technique with the potential to overcome these limitations.
• Assessing the performance of DLIR against established methods is crucial for its clinical adoption.

Purpose of the Study:

• To evaluate the physical performance of DLIR in comparison to FBP and IR.
• To quantify the image quality metrics, including noise and resolution, for each reconstruction method.
• To estimate the potential for radiation dose reduction using DLIR.

Main Methods:

• A cylindrical water phantom with embedded rods was scanned using a clinical CT scanner at varying dose levels (20, 15, 10, and 5 mGy).
• Images were reconstructed using FBP, DLIR, and IR.
• Task transfer functions (TTFs) and noise power spectrum (NPS) were measured to assess image quality and dose reduction potential.

Main Results:

• DLIR significantly reduced noise magnitude and preserved in-plane task transfer functions better than IR.
• DLIR demonstrated a higher estimated dose reduction potential (39%-54%) compared to IR (19%-29%).
• DLIR resulted in a slight decrease in axial resolution (6%-21%) compared to FBP.

Conclusions:

• DLIR exhibits superior in-plane edge-preserving noise reduction compared to IR.
• DLIR facilitates near 50% radiation dose reduction without compromising noise texture, provided some axial resolution reduction is acceptable.
• DLIR shows promise for dose-efficient CT imaging while maintaining diagnostic image quality.