Deep Learning-Based Versus Iterative Image Reconstruction for Unenhanced Brain CT: A Quantitative Comparison of Image...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Imaging

Background:

• Iterative reconstruction algorithms like Adaptive Iterative Dose Reduction 3D (AIDR-3D) are standard for CT imaging.
• Deep learning-based reconstruction, such as Automated Image Quality Enhancement (AiCE), is emerging as a potentially superior alternative.
• Quantitative comparison of these algorithms in unenhanced brain CT is crucial for clinical adoption.

Purpose of the Study:

• To quantitatively compare the image quality of unenhanced brain CT reconstructed using AIDR-3D and AiCE algorithms.
• To evaluate noise reduction, artifact levels, and contrast-to-noise ratios (CNRs) between the two reconstruction methods.
• To assess the performance variation of these algorithms in different brain areas.

Main Methods:

• Retrospective analysis of 100 unenhanced brain CT datasets acquired on a 320-detector row CT scanner.
• Comparison of AIDR-3D and AiCE reconstructions (0.5 mm thickness) using phantom and human studies.
• Calculation of image noise, artifact index in the posterior cranial fossa, and CNRs at cortical and thalamic levels.

Main Results:

• AiCE demonstrated significantly lower median image noise (19.6% reduction) and higher median CNRs at cortical and thalamic levels (p < 0.001).
• AIDR-3D exhibited a significantly lower artifact index in the posterior cranial fossa compared to AiCE (p < 0.001).
• Spatial resolution was comparable between the two algorithms in phantom studies.

Conclusions:

• Deep learning-based AiCE offers superior noise reduction and CNR in unenhanced brain CT compared to iterative AIDR-3D.
• Iterative AIDR-3D may provide better artifact management in specific regions like the posterior cranial fossa.
• The choice of reconstruction algorithm impacts image quality differently across various brain anatomical areas.