💥💥💞💞欢迎来到本博客❤️❤️💥💥
🏆博主优势:🌞🌞🌞博客内容尽量做到思维缜密,逻辑清晰,为了方便读者。
⛳️座右铭:行百里者,半于九十。
📋📋📋本文目录如下:🎁🎁🎁
目录
💥1 概述
基于监督学习的多模态MRI脑肿瘤分割利用监督体素的纹理特征。背景:由于各种肿瘤类型,精确分割磁共振成像(MRI)中的脑肿瘤是一项困难的任务。利用来自多模态MRI的信息和特征,包括结构MRI以及扩散张量成像(DTI)导出的各向同性(p)和各向异性(q)成分,可能会导致对脑图像的更准确的分析。<br>方法:我们提出了一种基于三维监督体素的学习方法,用于多模态MRI脑图像(常规MRI和DTI)中的肿瘤分割。监督体素是使用跨多模态MRI数据集的信息生成的。对于每个监督体素,提取了各种特征,包括使用一组不同大小和方向的Gabor滤波器计算的texton描述符的直方图,以及一阶强度统计特征。这些特征被输入到随机森林(RF)分类器中,以将每个监督体素分类为肿瘤核心、水肿或健康脑组织。<br>结果:该方法在两个数据集上进行了评估:1)我们的临床数据集:11名患者的多模态图像,2)BRATS 2013临床数据集:30个多模态图像。对于我们的临床数据集,使用多模态MRI检测肿瘤(包括肿瘤核心和水肿)的平均敏感性为86%,平衡错误率(BER)为7%;而自动肿瘤分割的Dice分数为0.84。BRATS 2013数据集的相应结果分别为96%,2%和0.89。结论:该方法在脑肿瘤分割中展现了有希望的结果。添加来自多模态MRI图像的特征可以大幅提高分割准确性。该方法在所有肿瘤级别上与专家标注基本匹配,从而实现了更快速和更可复制的脑肿瘤检测和标注方法,以帮助患者管理。
📚2 运行结果
部分代码:
% Initial arrays
I_temp = I(:,:,:,1);
S = nthroot((voxel_X * voxel_Y * voxel_Z),3);
X_window = floor(size(I_temp,1)/voxel_X);
Y_window = floor(size(I_temp,2)/voxel_Y);
Z_window = floor(size(I_temp,3)/voxel_Z);
r1 = ceil(voxel_X/2);
r2 = ceil(voxel_Y/2);
r3 = ceil(voxel_Z/2);
[X,Y,Z] = meshgrid((2:Y_window-1).*voxel_X-r2,(2:X_window-1).*voxel_Y-r1,(2:Z_window-1).*voxel_Z-r3);
% Initial Centers for 5 protocols
I_C = zeros([size(X),numel(ProtocolList)]);
for i=1:size(X,1)
for j = 1:size(X,2)
for l = 1:size(X,3)
I_C (i,j,l,P) = I(X(i,1,1), Y(1,j,1),Y(1,1,l),P);
end
end
end
C = [X(:),Y(:),Z(:)];
for P = 1:numel(ProtocolList)
I_C_temp = squeeze(I_C(:,:,:,P));
C = cat(2,C,I_C_temp(:));
end
k = size(C,1);
% Update supervoxel centers
for i = 1:k
temp_C = C(i,:);
x = temp_C(1);
y = temp_C(2);
z = temp_C(3);
temp_W = I_temp(x-1:x+1,y-1:y+1,z);
a = abs(gradient(temp_W));
[x,y] = find(a==min(a(:)), 1, 'first');
C(i,1) = temp_C(1)+x-2;
C(i,2) = temp_C(2)+y-2;
for P = 1:numel(ProtocolList)
C(i,3+1) = I(C(i,1), C(i,2),P);
end
end
% Assigning the lables
Label = -1*ones(size(I_temp));
Distance = inf*ones(size(I_temp));
counter = Iterations; % Iterastions
while counter
for c_k = 1:k
Center = [C(c_k,1),C(c_k,2),C(c_k,3)];
for i = Center(1)-2*r1+1:Center(1)+2*r1-1
for j = Center(2)-2*r2+1:Center(2)+2*r2-1
for l = Center(3)-2*r3+1:Center(3)+2*r3-1
% Calculating the spatial distance
d_c_term = 0;
d_c_temp = 0;
for P = 1:numel(ProtocolList)
d_c_temp = (I(Center(1),Center(2),Center(3),P)-I(i,j,l,P))^2;
d_c_term = d_c_term + d_c_temp;
end
d_c = sqrt(d_c_term );
% Calculating the intensity distance
d_s = sqrt((term_x*(Center(1)-i))^2 + (term_y*(Center(2)-j))^2+(term_z*(Center(3)-l))^2);
% Calculating the total distance D between C_k and i
D = sqrt((d_c)^2+(d_s/S)^2*Compactness^2);
if D < Distance(i,j,l)
Distance(i,j,l) = D;
Label(i,j,l) = c_k;
end
end % for l
end % for j
end % for i
end % for SV Label
%% Updating stage
counter = counter-1;
% Compute new cluster centers
if counter > 0
for c_k = 1:k
[x1,y1,z1] = find3((Label==c_k));
C(c_k,1) = round(mean(x1));
C(c_k,2) = round(mean(y1));
C(c_k,3) = round(mean(z1));
end
🎉3 参考文献
文章中一些内容引自网络,会注明出处或引用为参考文献,难免有未尽之处,如有不妥,请随时联系删除。
[1]Soltaninejad, Mohammadreza, et al. “Supervised Learning Based Multimodal MRI Brain Tumour Segmentation Using Texture Features from Supervoxels.” Computer Methods and Programs in Biomedicine, vol. 157, Elsevier BV, Apr. 2018, pp. 69–84, doi:10.1016/j.cmpb.2018.01.003.