Quantitative evaluation of a deep learning-based noise reduction algorithm in digital radiography using noise power...

Area of Science:

• Medical Imaging
• Digital Radiography
• Image Processing

Background:

• Deep learning (DL)-based noise reduction (DLNR) is increasingly used in clinical digital radiography (DR) systems.
• DLNR algorithms offer superior performance over conventional methods but function as "black boxes" with nonlinear behavior.
• Understanding the impact of DLNR on image quality across various imaging conditions is crucial.

Purpose of the Study:

• To quantitatively evaluate the image quality of a commercial DLNR algorithm, intelligent noise reduction (INR), in DR.
• To compare INR's noise reduction performance against a conventional rule-based algorithm (Con-NR).
• To utilize frequency-domain metrics, specifically noise power spectrum (NPS) analysis, for detailed comparison.

Main Methods:

• Noise power spectrum (NPS) analysis was employed to assess spatial-frequency-dependent noise characteristics.
• Both INR and Con-NR were evaluated across varying radiation dose levels and different imaging objects.
• A novel metric, the NPS improvement factor (NPSIF), was introduced to quantify noise suppression across frequency ranges.

Main Results:

• The DL-based INR algorithm demonstrated significant noise reduction at low radiation doses compared to Con-NR.
• The performance advantage of INR diminished at higher dose levels.
• The NPSIF effectively highlighted frequency-specific differences, revealing the distinct strengths and limitations of each noise reduction technique.

Conclusions:

• The dose-dependent performance of DL-based INR suggests its sensitivity to the training data characteristics.
• Distinct differences in noise suppression behavior exist between DL-based and conventional methods in DR.
• Frequency-domain analysis is essential for understanding advanced image processing techniques in DR; further research should integrate noise analysis with diagnostic performance.