RelativeFlow: Taming Medical Image Denoising Learning with Noisy Reference

2026-04-16 · Source: Artificial Intelligence · Field: Health & Wellbeing — Health & Medical Research, Medical Devices & Health Technology, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

RelativeFlow is a novel flow matching framework designed to address the noisy reference problem in medical image denoising (MID), where absolutely clean images for supervision are unavailable. Traditional methods like simulated-supervised discriminative learning (SimSDL) and simulated-supervised generative learning (SimSGL) often treat noisy references as clean targets, leading to suboptimal performance or reference-biased learning. Self-supervised learning (SSL) also struggles due to restrictive noise assumptions. RelativeFlow overcomes these limitations by learning from heterogeneous noisy references, driving inputs from various quality levels towards a unified high-quality target. It reformulates flow matching by decomposing the absolute noise-to-clean mapping into relative noisier-to-noisy mappings, implemented via consistent transport (CoT) and a simulation-based velocity field (SVF). Experiments on Computed Tomography (CT) and Magnetic Resonance (MR) denoising show RelativeFlow significantly outperforms existing methods.

Key takeaway

For AI Scientists developing medical image denoising solutions, RelativeFlow offers a robust approach to overcome the limitations of noisy reference data. Its novel flow matching framework, which learns relative noisier-to-noisy mappings, provides a path to achieve higher quality denoising without requiring perfectly clean ground truth. You should consider integrating this relative flow decomposition and consistent transport mechanism into your next-generation MID models, especially when working with diverse and imperfect medical imaging datasets.

Key insights

RelativeFlow improves medical image denoising by learning relative noise mappings from heterogeneous noisy references.

Principles

• Decompose absolute noise-to-clean mapping into relative mappings.
• Constrain relative flows to progressively compose a unified absolute flow.

Method

RelativeFlow uses consistent transport (CoT) to constrain relative flows and a simulation-based velocity field (SVF) with modality-specific degradation operators to construct a learnable velocity field.

In practice

• Apply RelativeFlow to CT denoising.
• Utilize RelativeFlow for MR denoising.

Topics

• RelativeFlow
• Medical Image Denoising
• Flow Matching
• Noisy Reference Problem
• Consistent Transport

Best for: AI Scientist, Research Scientist, Computer Vision Engineer

Related on AIssential

• RelativeFlow: Taming Medical Image Denoising Learning with Noisy Reference — RelativeFlow uses flow matching to learn unified denoising from heterogeneous noisy medical image references.
• Flow Map Denoisers: Traversing the Distortion-Perception Plane for Inverse Problems — Flow Map Denoisers use a single parameter to continuously balance distortion and perception in image restoration and inverse problems.
• The Third Challenge on Image Denoising at NTIRE 2026: Methods and Results — The NTIRE 2026 challenge benchmarks unconstrained neural architectures for high-noise image denoising, prioritizing PSNR performance.
• Plug-and-Play image restoration with Stochastic deNOising REgularization — SNORE improves PnP image restoration by applying denoisers to appropriately noisy images, enhancing performance and theoretical guarantees.
• FlowMaps: Modeling Long-Term Multimodal Object Dynamics with Flow Matching — FlowMaps uses latent flow matching to predict multimodal future object locations, improving robot navigation in dynamic 3D environments.
• Triangular Consistency as a Universal Constraint for Learning Optical Flow — Triangular consistency universally improves optical flow learning by enforcing geometric consistency among three composed flows, requiring no extra data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.