# Quantitative evaluation of a deep learning‐based noise reduction algorithm in digital radiography using noise power spectrum analysis

## TL;DR

This study evaluates a deep learning-based noise reduction algorithm in digital radiography using frequency-domain analysis to compare its performance with conventional methods.

## Contribution

The study introduces a new metric, the NPS improvement factor, to quantify frequency-specific noise suppression in deep learning-based noise reduction.

## Key findings

- The DL-based algorithm showed significant noise reduction at low-dose settings compared to conventional methods.
- The NPS improvement factor effectively captured frequency-specific differences between the algorithms.
- DL-based noise reduction performance was found to be dose-dependent and influenced by training data characteristics.

## Abstract

Recently, deep learning (DL)‐based noise reduction (DLNR) has been introduced in clinically used digital radiography (DR) systems, reporting superior performance over conventional algorithms. However, DLNR algorithms often operate as “black boxes” with nonlinear behavior, making it essential to understand the impact of such processing on image quality under different imaging conditions.

This study aimed to quantitatively evaluate the image quality of a commercial DLNR algorithm for DR referred to as intelligent noise reduction (INR). Specifically, we compared its noise reduction performance with that of a conventional rule‐based algorithm (conventional noise reduction, Con‐NR) using frequency‐domain metrics with detailed noise power spectrum (NPS) analysis.

The NPS was used to assess the spatial‐frequency‐dependent behavior of both INR and Con‐NR across varying dose levels and different objects. In this work, we introduced a supplementary metric—the NPS improvement factor (NPSIF)—to quantify noise suppression across frequency ranges and facilitate direct comparison between methods.

The DL‐based algorithm achieved substantial noise reduction at low‐dose settings compared with the conventional method, although its advantages were less pronounced at higher dose levels. The NPSIF effectively captured frequency‐specific differences, thereby offering insights into the strengths and limitations of each technique.

The dose‐dependent performance of the DL‐based algorithm suggests sensitivity to the characteristics of the training data used to develop the DL model. The findings demonstrate distinct differences in the noise suppression behavior between DL‐based and conventional methods in DR and underscore the importance of detailed frequency‐domain evaluation for understanding advanced image processing. Further research is warranted to integrate noise analysis with diagnostic performance metrics to comprehensively assess clinical utility.