Di-Flow: Self-Supervised Denoising for Single Volume MR Angiography via Slice-Aware Diffusion

SoJin Yun, Beomgu Kang, Sung-Hye You, Byungjun Kim, Dong-Hyun Kim, Hyunseok Seo

Keywords: Denoising, Self-Supervised Learning, Diffusion Models, MR Angiography

TL;DR: We propose a self-supervised diffusion-based denoising framework for single 3D MRA volumes that enhances anatomical structures without requiring clean labels or multi-view data.

Abstract:

Recent advances in deep learning have significantly improved medical image denoising, particularly through supervised convolutional neural network–based approaches. However, these rely on large-scale paired noisy–clean datasets, which limits their practical deployment in clinical settings where clean references are rarely available. Self-supervised methods mitigate this issue but typically depend on multi-volume data or temporal consistency, making them unsuitable for single 3D volume data like magnetic resonance angiography (MRA).
We propose a novel self-supervised denoising framework for single noisy volume that leverages spatial coherence across adjacent slices to construct training pairs without clean labels or repeated scans. At its core is a conditional denoising diffusion model with expectation-only sampling, enabling robust signal recovery. To further enhance anatomical fidelity, we introduce a patch-wise adaptive post-processing module that refines spatially localized features to better preserve anatomical accuracy. Validated on 7T and 3T time-of-flight MRA datasets, our method significantly improves vessel visibility while suppressing noise, offering a clinically practical denoising approach tailored to real-world imaging workflows.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 16205