三维大脑肿瘤分割使用深度学习

dataDir = fullfile (tempdir,“小鬼”);如果~存在(dataDir“dir”mkdir (dataDir);结束trained3DUnetURL =“//www.tianjin-qmedu.com/万博1manbetxsupportfiles/”+…“视觉/数据/ brainTumor3DUNetValid.mat”;downloadTrainedNetwork (trained3DUnetURL dataDir);负载(dataDir + filesep +“brainTumor3DUNetValid.mat”);

downloadBraTSSampleTestData (dataDir);

testDir = dataDir + filesep +“sampleBraTSTestSetValid”;data =加载(fullfile (testDir“imagesTest”,“BraTS446.mat”));标签=负载(fullfile (testDir“labelsTest”,“BraTS446.mat”));volt = data.cropVol;volTestLabels = labels.cropLabel;

bim = blockedImage (volt);

semanticsegBlock = @ (bstruct) semanticseg (bstruct.Data,净);

networkInputSize = net.Layers (1) .InputSize;networkOutputSize = net.Layers .OutputSize(结束);blockSize = [networkOutputSize (1:3) networkInputSize(结束)];borderSize = (networkInputSize (1:3)——blockSize (1:3)) / 2;

batchSize = 1;结果=应用(荡妇,…semanticsegBlock,…BlockSize = BlockSize,…BorderSize = BorderSize,…PadPartialBlocks = true,…BatchSize = BatchSize);predictedLabels = results.Source;

zID =大小(伏,3)/ 2;zSliceGT = labeloverlay(伏特(:,:,zID) volTestLabels (:,:, zID));zSlicePred = labeloverlay(伏特(:,:,zID) predictedLabels (:,:, zID));图蒙太奇({zSliceGT, zSlicePred},大小= [1 - 2],BorderSize = 5)标题(“贴上地面真理(左)与网络预测(右)”)

sourceDataLoc = dataDir + filesep +“Task01_BrainTumour”;preprocessDataLoc = dataDir + filesep +“preprocessedDataset”;preprocessBraTSDataset (preprocessDataLoc sourceDataLoc);

volLoc = fullfile (preprocessDataLoc,“imagesTr”);volds = imageDatastore (volLoc FileExtensions =“.mat”ReadFcn = @matRead);

lblLoc = fullfile (preprocessDataLoc,“labelsTr”);一会= [“背景”,“肿瘤”];pixelLabelID = [0 1];pixelLabelID pxds = pixelLabelDatastore (lblLoc,一会,…FileExtensions =“.mat”ReadFcn = @matRead);

patchSize = (132 132 132);patchPerImage = 16;miniBatchSize = 8;patchds = randomPatchExtractionDatastore (volds pxds patchSize,…PatchesPerImage = patchPerImage);patchds。MiniBatchSize = MiniBatchSize;

volLocVal = fullfile (preprocessDataLoc,“imagesVal”);voldsVal = imageDatastore (volLocVal FileExtensions =“.mat”,…ReadFcn = @matRead);lblLocVal = fullfile (preprocessDataLoc,“labelsVal”);pixelLabelID pxdsVal = pixelLabelDatastore (lblLocVal,一会,…FileExtensions =“.mat”ReadFcn = @matRead);dsVal = randomPatchExtractionDatastore (voldsVal pxdsVal patchSize,…PatchesPerImage = patchPerImage);dsVal。MiniBatchSize = MiniBatchSize;

numChannels = 4;inputPatchSize = [patchSize numChannels];numClasses = 2;[lgraph, outPatchSize] = unet3dLayers (inputPatchSize,…numClasses ConvolutionPadding =“有效的”);

dsTrain =变换(patchds,…@ (patchIn) augmentAndCrop3dPatch (patchIn outPatchSize,“培训”));dsVal =变换(dsVal,…@ (patchIn) augmentAndCrop3dPatch (patchIn outPatchSize,“确认”));

outputLayer = dicePixelClassificationLayer (Name =“输出”);lgraph = replaceLayer (lgraph,“Segmentation-Layer”,outputLayer);

inputLayer = image3dInputLayer (inputPatchSize,…归一化=“没有”、名称=“ImageInputLayer”);lgraph = replaceLayer (lgraph,“ImageInputLayer”,inputLayer);

deepNetworkDesigner (lgraph)

选择= trainingOptions (“亚当”,…MaxEpochs = 50,…InitialLearnRate = 5的军医,…LearnRateSchedule =“分段”,…LearnRateDropPeriod = 5,…LearnRateDropFactor = 0.95,…ValidationData = dsVal,…ValidationFrequency = 400,…情节=“训练进步”,…Verbose = false,…MiniBatchSize = MiniBatchSize);

doTraining =假;如果doTraining[网,信息]= trainNetwork (dsTrain、lgraph选项);modelDateTime =字符串(datetime (“现在”格式=“yyyy-MM-dd-HH-mm-ss”));保存(“trained3DUNet——”+ modelDateTime +“.mat”,“净”);结束

useFullTestSet =假;如果useFullTestSet volLocTest = fullfile (preprocessDataLoc,“imagesTest”);lblLocTest = fullfile (preprocessDataLoc,“labelsTest”);其他的volLocTest = fullfile (testDir,“imagesTest”);lblLocTest = fullfile (testDir,“labelsTest”);结束

voldsTest = imageDatastore (volLocTest FileExtensions =“.mat”,…ReadFcn = @matRead);pixelLabelID pxdsTest = pixelLabelDatastore (lblLocTest,一会,…FileExtensions =“.mat”ReadFcn = @matRead);

imageIdx = 1;datasetConfMat =表;而hasdata (voldsTest)%读和标签数据量卷=阅读(voldsTest);volLabels =阅读(pxdsTest);%为体积和创建blockedImage标签数据testVolume = blockedImage(卷);testLabels = blockedImage (volLabels {1});%计算块指标blockConfMatOneImage =应用(testVolume,…@(块,labeledBlock)…calculateBlockMetrics(块、labeledBlock网),…ExtraImages = testLabels,…PadPartialBlocks = true,…BlockSize = BlockSize,…BorderSize = BorderSize,…UseParallel = false);%阅读所有的块和更新的图像数量的结果blockConfMatOneImageDS = blockedImageDatastore (blockConfMatOneImage);blockConfMat = readall (blockConfMatOneImageDS);(blockConfMat blockConfMat = struct2table ({}):);blockConfMat。ImageNumber = imageIdx。*的(高度(blockConfMat), 1);datasetConfMat = [datasetConfMat; blockConfMat];imageIdx = imageIdx + 1;结束

[指标,blockMetrics] = evaluateSemanticSegmentation (…datasetConfMat,一会,指标=“所有”);

metrics.ImageMetrics.MeanIoU

函数blockMetrics = calculateBlockMetrics (bstruct gtBlockLabels净)%段块predBlockLabels = semanticseg (bstruct.Data,净);从gtBlockLabels %减少边境地区blockStart = bstruct。BorderSize + 1;blockEnd = blockStart + bstruct。BlockSize - 1;gtBlockLabels = gtBlockLabels (…blockStart (1): blockEnd (1),…blockStart (2): blockEnd (2),…blockStart (3): blockEnd (3));%评估分割结果对地面真理confusionMat = segmentationConfusionMatrix (predBlockLabels gtBlockLabels);% blockMetrics是混淆矩阵的结构,图像数量,%和块信息。blockMetrics。ConfusionMatrix = confusionMat;blockMetrics。ImageNumber = bstruct.ImageNumber;blockInfo。开始= bstruct.Start;blockInfo。结束= bstruct.End;blockMetrics。BlockInfo = BlockInfo;结束