RelativeFlow: Taming Medical Image Denoising Learning with Noisy Reference

A research team led by Yuting He and six other authors has developed RelativeFlow, a breakthrough AI framework that addresses a fundamental limitation in medical image denoising: the lack of absolutely clean reference images for training. Traditional approaches either treat noisy references as clean targets (causing suboptimal convergence) or impose restrictive noise assumptions that don't match real-world medical imaging scenarios. RelativeFlow solves this by reformulating flow matching to decompose the absolute noise-to-clean mapping into relative noisier-to-noisy mappings, allowing the system to learn from imperfect reference data.

The framework implements two key innovations: consistent transport (CoT), which constrains relative flows to progressively compose a unified absolute flow, and simulation-based velocity field (SVF), which constructs learnable velocity fields using modality-specific degradation operators. This enables RelativeFlow to handle different medical imaging modalities like CT and MRI while driving inputs from arbitrary quality levels toward a unified high-quality target. Extensive experiments demonstrate that RelativeFlow significantly outperforms existing simulated-supervised discriminative learning, simulated-supervised generative learning, and self-supervised learning methods.

The paper, accepted by CVPR 2026 and available on arXiv, represents a major advancement in medical AI by providing a practical solution to the noisy reference problem that has long limited denoising performance. By learning from heterogeneous noisy references rather than requiring perfect training data, RelativeFlow opens new possibilities for improving diagnostic imaging quality across healthcare settings where pristine reference images simply don't exist in practice.