Comparison of two deep learning image reconstruction algorithms in chest CT images: A task-based image quality...

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Healthcare

Background:

• Computed tomography (CT) imaging relies on image reconstruction algorithms.
• Deep learning reconstruction (DLR) offers potential improvements over traditional methods like filtered back-projection (FBP).
• Evaluating DLR algorithms is crucial for optimizing diagnostic performance in chest CT.

Purpose of the Study:

• To compare the performance of two distinct deep learning reconstruction (DLR) algorithms.
• To assess the impact of different DLR levels on image quality and lesion detectability in chest CT.
• To evaluate DLR performance across varying radiation dose levels and clinical indications.

Main Methods:

• Two CT scanners with different DLR algorithms (TrueFidelity™ and AiCE) were used.
• Image quality and phantom studies were conducted at six dose levels (0.5-10 mGy).
• Objective metrics (noise power spectrum, task-based transfer function, detectability index) and subjective radiologist assessments were performed.

Main Results:

• Both DLR algorithms reduced noise and improved lesion detectability compared to FBP.
• Noise magnitude varied between algorithms depending on the DLR level and dose.
• AiCE demonstrated higher detectability at high DLR levels, while TrueFidelity™ showed better spatial resolution in certain phantoms.

Conclusions:

• DLR algorithms effectively reduce image noise and enhance lesion detection in chest CT.
• The choice of DLR algorithm and reconstruction level impacts noise texture and spatial resolution.
• Clinical image quality was maintained across tested dose levels, with some variations based on algorithm and DLR level.