Fully automated image quality evaluation on patient CT: Multi-vendor and multi-reconstruction study

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Healthcare

Background:

• Computed tomography (CT) advancements aim to reduce radiation dose and image noise.
• Objective evaluation of CT image quality in patient scans remains a challenge.
• Standardized methods are needed for assessing image quality with new reconstruction algorithms.

Purpose of the Study:

• To present a patient-specific, automated method for evaluating CT image quality.
• To objectively measure noise level, structure sharpness, and structure alteration.
• To compare image quality metrics across different reconstruction algorithms, including deep learning.

Main Methods:

• Developed a method using structure coherence feature (SCF) to segment homogeneous (RH) and structure edge (RS) regions.
• Automated measurement of noise (standard deviation in RH ROIs) and structure sharpness (mean SCF in RS).
• Quantified structure alteration using standard deviation ratio on subtraction images between reconstruction methods.

Main Results:

• The automated method demonstrated high correlation (0.793) with manual edge slope measurements for structure sharpness.
• Iterative reconstruction (IR) and deep learning models (DLM) achieved 34.38% and 51.30% noise reduction compared to filtered back projection (FBP).
• DLM showed superior performance over IR in noise reduction, structure sharpness, and structure alteration, with significant noise reduction and minimal alteration.

Conclusions:

• The developed patient-specific method provides a high-throughput, quantitative evaluation of CT image quality.
• Deep learning models offer significant advantages in noise reduction and structure preservation compared to iterative reconstruction.
• This method can aid in optimizing CT protocols, especially for low-dose imaging practices.