RelativeFlow Framework Addresses Noisy Reference Problem in Medical Image Denoising

A novel framework known as RelativeFlow addresses the challenge of noisy references that hampers the effectiveness of medical image denoising. In contrast to current methods, RelativeFlow utilizes diverse noisy references and directs inputs of varying quality towards a common high-quality objective. This framework breaks down the absolute noise-to-clean relationship into relative mappings from noisier to noisy. It relies on two main elements: consistent transport (CoT), which serves as a displacement map, and an additional unspecified component. Traditional techniques, such as simulated-supervised discriminative learning (SimSDL) and simulated-supervised generative learning (SimSGL), incorrectly consider noisy references as clean targets, resulting in inadequate convergence. Self-supervised learning (SSL) often imposes unrealistic noise assumptions in practical medical imaging contexts. The findings were published on arXiv under identifier 2604.15459v1.