Reduction of radiation exposure in chest radiography using deep learning-based noise reduction processing: A phantom and retrospective clinical study

Abstract

Introduction

Intelligent noise reduction (INR), a deep learning-based noise reduction developed by Canon, is used in planar radiography to improve image quality and reduce patient exposure dose. This study aimed to evaluate the reduction of patient exposure dose in planar chest radiography using INR.

Methods

We evaluated the visibility of a Lungman phantom with tumor inserts by mean opinion score (MOS) to evaluate the optimal imaging conditions for INR. Furthermore, the optimal imaging conditions for INR were verified through retrospective evaluation using clinical images and the image quality was evaluated by blind/referenceless image spatial quality evaluator (BRISQUE). The individuals were the same 100 patients who had planar chest X-rays taken without INR and with INR, designated as the control and evaluation groups, respectively. Imaging conditions with automatic exposure control in the evaluation group set the radiation dose 32 % lower than that for the control group. The BRISQUE and entrance surface dose ($K_{a,e}$) in each group were compared.

Results

Regarding the visibility of the simulated mass, there was no significant difference in MOS when the reference dose was reduced by 33.33 % (p = 0.26). In retrospective evaluation of clinical images, BRISQUE in the control and evaluation groups was 34.35 ± 4.19 and 34.46 ± 4.58 (p = 0.35), respectively. The $K_{a,e}$ in the control and evaluation groups were 0.131 ± 0.039 and 0.084 ± 0.024 mGy (p < 0.001).

Conclusion

INR reduced patient exposure dose by an average of 35 % without decreasing image quality.

Implications for practice

These results indicate that INR can contribute to the reduction of patient radiation dose during chest radiography. The widespread use of this technology may reduce dose indices, including diagnostic reference levels.