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

This study aimed to assess the image characteristics of deep-learning-based image processing software (DLIP; FCT PixelShine, FUJIFILM, Tokyo, Japan) and compare it with filtered back projection (FBP), model-based iterative reconstruction (MBIR), and deep-learning-based reconstruction (DLR).

This phantom study assessed the object-specific spatial resolution (task-based transfer function [TTF]), noise characteristics (noise power spectrum [NPS]), and low-contrast detectability (low-contrast object-specific contrast-to-noise ratio [CNR_LO]) at three different output doses (standard: 10 mGy; low: 3.9 mGy; ultralow: 2.0 mGy). The processing strength of DLIP_FBP with A1, A4, and A9 was compared with those of FBP, MBIR, and DLR.

The standard dose with high-contrast TTFs of DLIP_FBP exceeded that of FBP. Low-contrast TTFs were comparable to or lower than that of FBP. The NPS peak frequency (f_P) of DLIP_FBP shifts to low spatial frequencies of up to 8.6% at ultralow doses compared to the standard FBP dose. MBIR shifted the most f_P compared to FBP-a marked shift of up to 49%. DLIP_FBP showed a CNR_LO equal to or greater than that of DLR in standard or low doses. In contrast, the CNR_LO of the DLIP_FBP was equal to or lower than that of the DLR in ultralow doses.

DLIP_FBP reduced image noise while maintaining a resolution similar to commercially available MBIR and DLR. The slight spatial frequency shift of f_P in DLIP_FBP contributed to the noise texture degradation suppression. The NPS suppression in the low spatial frequency range effectively improved the low-contrast detectability.