Diffusion probabilistic versus generative adversarial models to reduce contrast agent dose in breast MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Reducing contrast agent dose in MRI-based breast cancer screening is desirable.
• Deep learning models can potentially enhance signal in low-dose contrast-enhanced breast MRI.

Purpose of the Study:

• To compare denoising diffusion probabilistic models (DDPM) and generative adversarial networks (GAN) for reconstructing low-dose contrast-enhanced breast MRI subtraction images.
• To evaluate the performance of DDPM and GAN at different dose reduction levels (25%, 10%, 5%).

Main Methods:

• Retrospective analysis of 50 breasts with enhancing lesions and 50 lesion-free examinations.
• Comparison of DDPM- and GAN-reconstructed images against original images using quantitative measures and radiologic evaluations.
• Radiologist preference assessment for reconstruction quality and lesion conspicuity, and evaluation of false-positive rates.

Main Results:

• Radiologists preferred GAN images at 5% dose and DDPM images at 25% dose.
• Lesion conspicuity scores were similar between GAN and DDPM at 25% and 10% doses, but higher for GAN at 5% dose.
• No significant difference in false-positive rates between DDPM and GAN across all tested dose levels.

Conclusions:

• Both GAN and DDPM show potential for low-dose contrast-enhanced breast MRI reconstruction.
• Neither model demonstrated superiority across all dose levels and evaluation metrics.
• Further research is needed to mitigate false-positives and optimize model performance.