Impact of an artificial intelligence deep-learning reconstruction algorithm for CT on image quality and potential...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep-learning reconstruction algorithms are emerging in CT to address limitations of iterative reconstruction (IR).
• These new algorithms aim to overcome issues like image smoothing and resolution dependency on contrast and dose.
• This study evaluates a specific AI-DLR algorithm against traditional IR methods.

Purpose of the Study:

• To evaluate the impact of an artificial intelligence deep-learning reconstruction (AI-DLR) algorithm on image quality in chest CT.
• To compare AI-DLR with hybrid iterative reconstruction (IR) regarding image quality and potential dose reduction.
• To assess performance across various clinical indications in chest CT.

Main Methods:

• Chest CT phantom studies (ACR 464, Torso CTU-41) were conducted at five dose levels.
• Raw data were reconstructed using filtered backprojection, IR (iDose4 levels 4 and 7), and AI-DLR (Smoother to Sharper levels).
• Image quality was assessed using noise power spectrum, transfer function, detectability index (d') for lesions (nodules, GGO, high-contrast lesions), and subjective radiologist evaluation.

Main Results:

• Smoother AI-DLR levels significantly decreased noise (-63% to -66%) and NPS spatial frequency, while increasing lesion detectability (d').
• Sharper AI-DLR levels showed the opposite trend, increasing spatial resolution for low-contrast inserts but decreasing it for high-contrast inserts.
• AI-DLR levels (Smoother, Smooth, Standard) improved detectability compared to clinical IR (i4), with radiologists deeming images satisfactory but noting the need for dose adaptation.

Conclusions:

• AI-DLR effectively reduces noise and enhances lesion detection at its smoother settings, albeit with increased image smoothing.
• Conversely, the sharper AI-DLR settings improve spatial resolution but may reduce noise reduction and lesion detectability.
• The optimal AI-DLR level selection is crucial and should be tailored based on the specific radiation dose and whether mediastinal or parenchymal images are being analyzed.