Physical characteristics of deep learning-based image processing software in computed tomography: a phantom study.

Area of Science:

• Medical imaging
• Radiology
• Image processing

Background:

• Filtered back projection (FBP) is a traditional image reconstruction method.
• Model-based iterative reconstruction (MBIR) and deep-learning-based reconstruction (DLR) offer advanced image quality.
• Deep-learning-based image processing (DLIP) software presents a novel approach to enhance medical images.

Purpose of the Study:

• To evaluate the image characteristics of DLIP software (FCT PixelShine).
• To compare DLIP performance against FBP, MBIR, and DLR.
• To assess image quality metrics including spatial resolution, noise, and low-contrast detectability.

Main Methods:

• A phantom study was conducted to assess object-specific spatial resolution (task-based transfer function [TTF]), noise power spectrum (NPS), and low-contrast detectability (contrast-to-noise ratio [CNR_LO]).
• Evaluations were performed at standard (10 mGy), low (3.9 mGy), and ultralow (2.0 mGy) radiation doses.
• The processing strength of DLIP_FBP was compared with FBP, MBIR, and DLR.

Main Results:

• DLIP_FBP demonstrated superior high-contrast TTFs compared to FBP at standard doses.
• Low-contrast TTFs with DLIP_FBP were comparable or lower than FBP.
• DLIP_FBP shifted NPS peak frequency to lower spatial frequencies, especially at ultralow doses, unlike MBIR which showed a more significant shift.
• DLIP_FBP achieved CNR_LO equal to or greater than DLR at standard and low doses, but lower at ultralow doses.

Conclusions:

• DLIP_FBP effectively reduces image noise while preserving spatial resolution comparable to MBIR and DLR.
• The observed shift in spatial frequency (f_P) in DLIP_FBP aids in suppressing noise texture degradation.
• Suppression of NPS in the low spatial frequency range significantly enhances low-contrast detectability.