基于MATLAB的MRI图像去噪实现

基于MATLAB的MRI图像去噪实现，整合了非局部均值滤波（NLM）、BM3D算法及改进的ADMM-TGV方法

一、非局部均值滤波（NLM）实现

functiondenoised=nlm_mri(noisy,h=0.1,patch_size=5,search_size=13)% 参数说明：% noisy: 输入噪声MRI图像 (三维矩阵)% h: 滤波参数（控制平滑强度）% patch_size: 相似块大小% search_size: 搜索窗口大小[X,Y,Z]=size(noisy);denoised=noisy;pad=floor(patch_size/2)+search_size//2;noisy_pad=padarray(noisy,[pad pad pad],'symmetric');fori=1+pad:X+padforj=1+pad:Y+padfork=1+pad:Z+pad% 定义搜索窗口center=[i,j,k];window=get_window(center,search_size,X,Y,Z);% 提取相似块patches=cell(size(window));forp=1:numel(window)patches{p}=get_patch(noisy_pad,window{p},patch_size);end% 计算权重矩阵weights=zeros(size(patches));forp=1:numel(patches)diff=patches{p}-mean(patches{p}(:));weights(:,:,p)=exp(-sum(diff.^2,3)/(h^2));end% 加权平均total_weight=sum(weights(:));iftotal_weight>0denoised(i-pad,j-pad,k-pad)=...sum(sum(sum(patches.*weights)))/total_weight;endendendendend

二、三维BM3D算法实现

functiondenoised=bm3d_mri(noisy,sigma=0.1,block_size=8,search_range=16)% 参数说明：% noisy: 输入噪声MRI图像 (三维矩阵)% sigma: 噪声标准差估计% block_size: 分块尺寸% search_range: 块匹配搜索范围[X,Y,Z]=size(noisy);denoised=noisy;padded=padarray(noisy,[block_size block_size block_size],'symmetric');% 分块处理fori=1:block_size:X+block_size-1forj=1:block_size:Y+block_size-1fork=1:block_size:Z+block_size-1% 提取当前块current_block=padded(i:i+block_size-1,j:j+block_size-1,k:k+block_size-1);% 块匹配matches=find_similar_blocks(padded,current_block,search_range);% 三维变换与阈值处理group=cat(4,current_block,matches);coeffs=dct3(group);threshold=sigma*sqrt(2*log(size(coeffs,4)));coeffs(abs(coeffs)<threshold)=0;% 逆变换与加权平均denoised_block=idct3(coeffs);denoised(i:i+block_size-1,j:j+block_size-1,k:k+block_size-1)=...mean(denoised_block,4);endendendend

三、改进的ADMM-TGV算法

functiondenoised=admm_tgv_mri(noisy,lambda=0.1,rho=1,max_iter=100)% 参数说明：% noisy: 输入噪声MRI图像 (三维矩阵)% lambda: 正则化参数% rho: ADMM惩罚参数% max_iter: 最大迭代次数[X,Y,Z]=size(noisy);denoised=noisy;z=zeros(size(noisy));u=zeros(size(noisy));foriter=1:max_iter% 数据保真项更新residual=noisy-z+u;z_prev=z;z=residual+(lambda/rho)*TV_denoise(residual);% 总广义变分正则项更新grad_x=z(:,:,2:end)-z(:,:,1:end-1);grad_y=z(:,2:end,:)-z(:,1:end-1,:);grad_z=z(2:end,:,:)-z(1:end-1,:,:);dual=cat(4,grad_x,grad_y,grad_z);u=prox_l1(dual,rho);% 收敛判断ifnorm(z-z_prev)/norm(z_prev)<1e-3break;endendendfunctionout=TV_denoise(in)% Total Variation去噪[X,Y,Z]=size(in);out=in;fori=2:X-1forj=2:Y-1fork=2:Z-1out(i,j,k)=median(in(i-1:i+1,j-1:j+1,k-1:k+1),'all');endendendend

四、性能对比与参数调优

参数调优建议：

1. BM3D：block_size=8适用于1.5T MRI，sigma通过Rician噪声估计获取
2. ADMM-TGV：lambda取0.05-0.2范围，rho建议1-5之间调整
3. NLM：search_size=13平衡计算效率与去噪效果

五、完整应用示例

%% 加载MRI数据img=dicomread('brain.dcm');noisy=imnoise(img,'gaussian',0,0.01);%% 去噪处理tic;denoised_bm3d=bm3d_mri(noisy,0.12,8,16);time_bm3d=toc;tic;denoised_admm=admm_tgv_mri(noisy,0.15,2,150);time_admm=toc;%% 结果可视化figure;subplot(1,3,1);imshow(squeeze(img(:,:,50)),[]);title('原始图像');subplot(1,3,2);imshow(squeeze(denoised_bm3d(:,:,50)),[]);title('BM3D去噪');subplot(1,3,3);imshow(squeeze(denoised_admm(:,:,50)),[]);title('ADMM-TGV去噪');%% 性能评估psnr_bm3d=psnr(denoised_bm3d,img);ssim_bm3d=ssim(denoised_bm3d,img);psnr_admm=psnr(denoised_admm,img);ssim_admm=ssim(denoised_admm,img);disp(['BM3D - PSNR=%.2f dB, SSIM=%.4f, 耗时=%.1fs',psnr_bm3d,ssim_bm3d,time_bm3d]);disp(['ADMM-TGV - PSNR=%.2f dB, SSIM=%.4f, 耗时=%.1fs',psnr_admm,ssim_admm,time_admm]);

六、工程实践建议

1. 预处理流程： 格式转换：DICOM → NIfTI（使用dcm2niix工具） 头骨去除：采用FSL的BET工具 归一化：将信号强度缩放到[0,1]区间

2. 并行加速：

   % 启用并行计算池parpool('local',4);% 并行化BM3D块处理parfori=1:block_size:X+block_size-1% 块处理代码end

3. GPU加速：

   % 将数据转移至GPUnoisy_gpu=gpuArray(noisy);% 在GPU上执行计算denoised_gpu=bm3d_mri(noisy_gpu);denoised=gather(denoised_gpu);

七、参考

1. 代码: 用于进行MRI图像去噪的源码www.youwenfan.com/contentcsp/98348.html
2. 工具包：DIPY：扩散MRI处理工具（含Patch2Self）SPM12：标准MRI预处理流程TensorFlow Addons：提供MRI专用层