DiffDenoise: Self-Supervised Medical Image Denoising with Conditional Diffusion Models

2025-03-31

AI Summary

Existing self-supervised medical image denoising methods often over-smooth images, leading to loss of critical anatomical details. To address this, we propose a three-stage, ground-truth-free, high-fidelity denoising framework: (1) A diffusion model is conditioned on blind-spot network (BSN) outputs to enhance high-frequency modeling; (2) a symmetric noise averaging strategy is introduced to stabilize the reverse sampling process; and (3) knowledge distillation generates high-quality pseudo-labels, enabling joint optimization of self-supervised and semi-supervised objectives. Our method is the first to integrate conditional diffusion modeling with symmetric noise stabilization for medical denoising. Evaluated on diverse real and synthetic medical modalities—including CT, MRI, and ultrasound—it achieves significant improvements over state-of-the-art methods in PSNR and SSIM, demonstrates strong cross-modality generalization, and maintains theoretical soundness.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Unsupervised & Self-Supervised Learning
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

Many self-supervised denoising approaches have been proposed in recent years. However, these methods tend to overly smooth images, resulting in the loss of fine structures that are essential for medical applications. In this paper, we propose DiffDenoise, a powerful self-supervised denoising approach tailored for medical images, designed to preserve high-frequency details. Our approach comprises three stages. First, we train a diffusion model on noisy images, using the outputs of a pretrained Blind-Spot Network as conditioning inputs. Next, we introduce a novel stabilized reverse sampling technique, which generates clean images by averaging diffusion sampling outputs initialized with a pair of symmetric noises. Finally, we train a supervised denoising network using noisy images paired with the denoised outputs generated by the diffusion model. Our results demonstrate that DiffDenoise outperforms existing state-of-the-art methods in both synthetic and real-world medical image denoising tasks. We provide both a theoretical foundation and practical insights, demonstrating the method's effectiveness across various medical imaging modalities and anatomical structures.

Problem

• Self-supervised medical image denoising preserving fine structures
• Overcoming over-smoothing in existing denoising methods
• Enhancing high-frequency detail retention in medical images

Innovation

• Self-supervised denoising with conditional diffusion models
• Stabilized reverse sampling using symmetric noises
• Supervised training with diffusion-generated clean images