DiffDenoise: Conditional Diffusion Transforms Medical Image Denoising

Introduction

Medical imaging is essential for disease diagnosis, but noise often degrades image quality. DiffDenoise is a self‑supervised three‑stage pipeline that combines a Blind‑Spot Network (BSN) with a conditional diffusion model and Stabilized Reverse Diffusion Sampling (SRDS) to achieve high‑quality denoising on both synthetic and real medical datasets.

Key Terminology

BSN (Blind‑Spot Network) : predicts the value of a center pixel from its surrounding pixels while masking the center, enabling self‑supervised training.

SRDS (Stabilized Reverse Diffusion Sampling) : generates two diffusion samples with opposite (positive and negative) noise and averages them to cancel sampling bias.

DDIM (Denoising Diffusion Implicit Models) : a non‑Markovian diffusion sampling technique that accelerates inference.

Method Overview

Stage 1 – Pre‑train BSN : The BSN is trained on noisy images in a self‑supervised manner. Its output is a “half‑finished” denoised image that reduces the noise standard deviation by roughly threefold.

Stage 2 – Conditional Diffusion : The BSN output is used as a conditioning signal for a diffusion model. The diffusion model learns to map the noisy image (plus the BSN condition) to a clean image by predicting the added noise at each diffusion step.

Stage 3 – Stable Sampling + Knowledge Distillation : During inference, SRDS initializes the reverse diffusion process with both positive and negative noise, runs the diffusion steps, and averages the two results to lower variance. A lightweight student network is then trained to mimic the diffusion output, providing fast inference.

Core Design Details

Symmetric Noise Trick : By initializing the reverse diffusion with both +ε and –ε and averaging the final samples, the variance of the estimator drops from 0.7 to 0.3, yielding a 2–3 dB PSNR gain.

Knowledge Distillation : A compact student network (≈0.5 s per image) is trained to reproduce the diffusion model’s output (≈5 s for 20 DDIM steps). The PSNR loss is only ~0.1 dB, giving a ten‑fold speedup.

BSN as Condition : Although the BSN discards some fine details, its output reduces the noise standard deviation by a factor of three, providing a strong prior that lets the diffusion model focus on the remaining ~30 % of details.

Datasets & Experimental Setup

Synthetic datasets : FastMRI and COVID‑19 chest X‑ray images corrupted with four noise types:

Gaussian (σ = 6/255)

Poisson (λ = 200/700)

Gamma (α = 100/500, β = 100/500)

Spatially correlated noise (σ = 0.5/1.2)

Real datasets : M4Raw MRI T1‑weighted and FLAIR sequences. Clean references are obtained by averaging multiple scans of the same subject.

Results

On a knee‑joint MRI test set, DiffDenoise improves PSNR from 32.57 dB (baseline) to 36.02 dB. The method excels on spatially correlated noise, maintaining >35 dB where other methods fall below 35 dB. On the real‑world M4Raw MRI set, DiffDenoise reaches 31.95 dB PSNR, surpassing all self‑supervised baselines and approaching the supervised NAFNet result of 32.17 dB.

Analysis

SRDS reduces sampling variance by 58 % compared with two independent random samples, directly translating into the observed 2–3 dB PSNR gain. Knowledge distillation retains almost all performance while delivering a ten‑fold inference speedup (0.5 s vs. 5 s). The BSN‑conditioned diffusion remains robust when the BSN blind‑spot radius is enlarged: PSNR drops only 0.5 dB versus a 2 dB drop for a pure BSN pipeline.

Conclusion & Future Work

DiffDenoise establishes a new benchmark for medical image denoising by integrating conditional diffusion, symmetric sampling, and knowledge distillation. Future extensions may apply the same three‑stage framework to super‑resolution, artifact removal, and broader clinical deployment.

Code example

收
藏
，
分
享
、
在
看
，
给
个
三
连
击呗！