RelativeFlow: Taming Medical Image Denoising Learning with Noisy Reference
Summary: arXiv:2604.15459v1 Announce Type: cross
Abstract
Medical image denoising (MID) is a critical task that faces significant challenges due to the lack of absolutely clean images for supervision. This situation leads to a noisy reference problem that fundamentally limits the performance of denoising algorithms. Existing methods, such as simulated-supervised discriminative learning (SimSDL) and simulated-supervised generative learning (SimSGL), treat noisy references as clean targets. This approach often results in suboptimal convergence and reference-biased learning. On the other hand, self-supervised learning (SSL) imposes strict noise assumptions that are seldom satisfied in realistic MID scenarios.
Introduction to RelativeFlow
In response to these challenges, we propose RelativeFlow, a novel flow matching framework designed to learn from heterogeneous noisy references. This innovative approach aims to drive inputs from various quality levels toward a unified high-quality target. RelativeFlow’s methodical reformulation of flow matching decomposes the absolute noise-to-clean mapping into relative noisier-to-noisy mappings. This is realized through two key components:
- Consistent Transport (CoT): A displacement map that constrains relative flows to be components of and progressively compose a unified absolute flow.
- Simulation-based Velocity Field (SVF): Constructs a learnable velocity field using modality-specific degradation operators to support various medical imaging modalities.
Methodology
The RelativeFlow framework operates by utilizing the strengths of both consistent transport and simulation-based velocity fields. The CoT ensures that the learning process maintains consistency across various noisy references, while the SVF adapts to the specific degradation characteristics of different imaging modalities, such as Computed Tomography (CT) and Magnetic Resonance (MR).
Experimental Results
Extensive experiments have been conducted to evaluate the performance of RelativeFlow in the context of CT and MR denoising. The results demonstrate that RelativeFlow significantly outperforms existing methods in effectively taming medical image denoising with noisy references. This advancement not only enhances the quality of medical images but also broadens the applicability of denoising techniques in clinical settings.
Conclusion
RelativeFlow represents a significant step forward in the field of medical image denoising. By addressing the noisy reference problem and leveraging heterogeneous noisy images, this framework provides a robust solution that improves denoising performance across various modalities. Future work will focus on refining the framework further and exploring its potential applications in other imaging domains.
Key Takeaways
- RelativeFlow addresses the challenges of medical image denoising due to lack of clean references.
- The framework introduces novel components such as Consistent Transport and Simulation-based Velocity Field.
- Extensive experiments validate the superior performance of RelativeFlow in CT and MR denoising tasks.