CNN 脑肿瘤诊断方案MATLAB 实现

CNN 脑肿瘤诊断方案MATLAB 实现

---

一、CNN 在脑肿瘤诊断中的两种典型用法（MATLAB ）

任务	输入	输出	常用 CNN
分类	2D MRI 切片	肿瘤类型 / 是否患病	ResNet / AlexNet / GoogLeNet
分割	2D / 3D MRI	肿瘤区域掩码	U‑Net / SegNet

MATLAB 官方强烈推荐：分类用 ResNet，分割用 U‑Net

---

二、脑肿瘤分类（ResNet18，最稳妥方案）

1、数据集准备（标准结构）

dataset/
 ├── glioma/
 ├── meningioma/
 ├── pituitary/
 └── normal/

每张图片为 jpg / png / bmp

---

2、数据加载与增强（关键）

dataDir = "dataset";

imds = imageDatastore( ...
    dataDir, ...
    "IncludeSubfolders", true, ...
    "LabelSource", "foldernames");

% 数据增强（医学影像必备）
aug = imageDataAugmenter( ...
    "RandRotation",[-15 15], ...
    "RandXReflection",true, ...
    "RandYReflection",true);

auimds = augmentedImageDatastore([224 224], imds, ...
    "DataAugmentation", aug);

---

3、构建 CNN（迁移学习）

net = resnet18;

lgraph = layerGraph(net);

% 替换全连接层
newFc = fullyConnectedLayer( ...
    4, ...
    "Name","fc_new", ...
    "WeightLearnRateFactor",10, ...
    "BiasLearnRateFactor",10);

lgraph = replaceLayer(lgraph,"fc1000",newFc);
lgraph = replaceLayer(lgraph,"prob", ...
    softmaxLayer("Name","softmax"));
lgraph = replaceLayer(lgraph,"ClassificationLayer_predictions", ...
    classificationLayer("Name","output"));

---

4、 训练 CNN

options = trainingOptions("adam", ...
    "InitialLearnRate",1e-4, ...
    "MaxEpochs",15, ...
    "MiniBatchSize",16, ...
    "Shuffle","every-epoch", ...
    "ValidationData",auimds, ...
    "ValidationFrequency",30, ...
    "Plots","training-progress", ...
    "Verbose",false);

[netTrained, info] = trainNetwork(auimds, lgraph, options);

---

5、 测试与评估

testImds = imageDatastore( ...
    "test_dataset", ...
    "IncludeSubfolders",true, ...
    "LabelSource","foldernames");

YPred = classify(netTrained, testImds);
YTrue = testImds.Labels;

accuracy = mean(YPred == YTrue)

confusionchart(YTrue, YPred)

典型结果：
4 类脑肿瘤准确率 94%–97%（BraTS / Figshare 数据）

---

三、脑肿瘤分割（U‑Net，临床级方案）

1、数据格式（分割）

文件	内容
images/	MRI 切片
masks/	肿瘤掩码（0/1/2/3）

---

2、构建 U‑Net

imageSize = [256 256 1];

encoderDepth = 3;
numClasses = 4;

unet = unetLayers(imageSize, numClasses, ...
    "EncoderDepth", encoderDepth);

---

3、训练 U‑Net

options = trainingOptions("adam", ...
    "InitialLearnRate",1e-3, ...
    "MaxEpochs",20, ...
    "MiniBatchSize",8, ...
    "Shuffle","every-epoch", ...
    "Plots","training-progress");

[netSeg, info] = trainNetwork(dsTrain, unet, options);

---

4、分割结果可视化

[YPred, scores] = semanticseg(testImg, netSeg);

figure
montage({testImg, YPred})
title("MRI 与 肿瘤分割结果")

参考代码 CNN卷积神经网络在脑肿瘤诊断中的应用 www.youwenfan.com/contentcsu/63411.html

四、评价指标

分类

• Accuracy
• Precision / Recall
• F1‑score
• Confusion Matrix

分割

• Dice Coefficient
• IoU
• Sensitivity / Specificity

dice = diceOverlap(YTrue, YPred);

---

五、工程级建议

1. 2D vs 3D

场景	建议
毕业设计	2D CNN
论文 / 项目	3D CNN（volumetric）
实时系统	2D + 滑动窗口

2. 多模态 MRI

同时输入：

• T1
• T2
• FLAIR

通道数 = 3（像 RGB）