The INR algorithms using deep-learning demonstrated particularly notable improvements in NPS under low-dose conditions, while its effect tended to be smaller under high-dose conditions. This behavior is considered to mean that the performance of deep learning model depends on the training images, suggesting that the noise characteristics inherent in the training data define the upper limit of noise improvement.

NPSIF, which quantitatively evaluates the effects of image processing for each spatial frequency, offers a more detailed and intuitive characterization of noise reduction techniques compared to traditional evaluations using only NPS; this approach is valuable in clearly identifying subtle differences between processing methods. Specifically, it played an essential role in comparing the improvement effects of conventional and INR processing, aiding in the understanding of frequency dependence and its limitations. Therefore, the findings of this study provide important insights for selecting and implementing noise reduction techniques in clinical practice. The results suggest that NPSIF can be utilized to identify the optimal image processing method tailored to varying dose conditions and patients.

While NPSIF is effective for evaluating the detailed characteristics of noise reduction techniques, it may need to be combined with other evaluation metrics. For instance, incorporating measures related to diagnostic performance, such as spatial resolution and contrast, enables a more comprehensive analysis. We believe that this can be overcome by evaluating a task-based SNR improvement factor that combines the NPSIF and the MTF improvement effect, focusing on the effects of image processing for specific diagnostic tasks.

This study demonstrated that the newly proposed NPSIF is an effective metric for quantitatively evaluating the operational characteristics of noise reduction techniques across spatial frequencies. This metric clarifies the properties and limitations of different techniques that conventional methods fail to capture, contributing to the selection of appropriate algorithms in clinical practice. The findings from this study provide important scientific evidence for imaging technologists to reduce patient exposure while delivering high-quality diagnostic images.