Digital radiography (DR) is a widely used in clinical practice. The exposure dose per examination in DR is lower compared to other modalities, however, it is one of the factors contributing to integrated exposure of patients [1,2]. Therefore, reducing the exposure dose while maintaining diagnostic image quality of DR is a crucial task for technologists [3–6].

The reduction of the imaging dose leads to an increase in quantum noise. Since image noise is an important factor affecting the accuracy of diagnostic imaging [7], various approaches to reduce noise have been implemented [8–10]. In particular, recent integration of deep-learning technology into radiology has shown its effectiveness, both in improving image quality and in establishing dose-reduction protocols [11,12].

To effectively utilize such cutting-edge technology and provide benefits to patients, it is essential to clearly understand their operational characteristics. Many of the newer deep-learning-based image processess exhibit much more complex behavior compared to traditional algorithms. Furthermore, the internal mechanisms of these models are essentially black boxes, making it difficult to predict their outputs under various input conditions [13]. Nevertheless, the nonlinear behavior resulting from their complex processes influences the spatial correlation of the image during noise reduction, so it is necessary to grasp these effects [14]. 

In this study, we aimed to address these issues by proposing a novel evaluation metric called the Noise Power Spectrum (NPS) Improvement Factor (NPSIF) and verifying its effectiveness. This metric enables detailed quantitative evaluation of the noise reduction effect of image processing based on spatial frequency characteristics.