Generative adversarial network with radiomic feature reproducibility analysis for computed tomography denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiomics

Background:

• Traditional image quality analysis (IQA) methods for computed tomography (CT) are insufficient for medical imaging texture details.
• Radiomics offers objective texture analysis for medical diagnostics, overcoming subjective limitations.
• Evaluating CT denoising requires metrics that account for imaging protocol variations.

Purpose of the Study:

• Introduce radiomic feature reproducibility analysis as a novel evaluation metric for CT denoising algorithms.
• Propose a low-dose CT denoising method utilizing a generative adversarial network (GAN).
• Compare the proposed method's performance against conventional CT denoising techniques.

Main Methods:

• Implemented radiomic feature reproducibility analysis to assess CT denoising performance.
• Developed a generative adversarial network (GAN)-based denoising algorithm for low-dose CT.
• Utilized radiomic feature reproducibility for hyper-parameter tuning of the GAN.

Main Results:

• Traditional metrics (PSNR, SSIM) did not distinguish the proposed GAN denoising method from conventional ones.
• Radiomic feature reproducibility analysis effectively differentiated the performance of the CT denoising methods.
• The GAN-based denoising method, fine-tuned using radiomics, demonstrated superior performance.

Conclusions:

• The proposed GAN architecture, optimized with radiomics, outperforms existing CT denoising methods.
• This study pioneers the use of radiomics reproducibility analysis for evaluating CT denoising.
• The findings aim to bridge objective and subjective evaluation gaps in clinical medical imaging.