brain tumor dataset

姜智浩 Lv5
  2025-12-02 11:45:14 2025-12-02 11:45:14 Created   2025-12-02 10:36:41 2025-12-02 10:36:41 Updated      1.1k Words  6 Mins  

数据来源

https://www.kaggle.com/datasets/masoudnickparvar/brain-tumor-mri-dataset

导入库

1
2
3
4
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
from tensorflow.keras.applications.densenet import preprocess_input

检查是否启用了GPU

1
print(tf.config.list_physical_devices("GPU"))

[PhysicalDevice(name=’/physical_device:GPU:0’, device_type=’GPU’)]

构造数据集

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
# 参数
BATCH_SIZE = 32
IMG_SIZE = (224, 224)
DATA_DIR = "/Volumes/HIKSEMI/ipynb/Kaggle/Brain Tumor Dataset"

train_ds_raw = tf.keras.preprocessing.image_dataset_from_directory(
    DATA_DIR,
    validation_split=0.2,
    subset="training",
    seed=123,
    color_mode="rgb",
    image_size=IMG_SIZE,
    batch_size=BATCH_SIZE,
)
val_ds_raw = tf.keras.preprocessing.image_dataset_from_directory(
    DATA_DIR,
    validation_split=0.2,
    subset="validation",
    seed=123,
    color_mode="rgb",
    image_size=IMG_SIZE,
    batch_size=BATCH_SIZE,
)

Found 10560 files belonging to 4 classes.
Using 8448 files for training.
Found 10560 files belonging to 4 classes.
Using 2112 files for validation.

1
2
3
4
5
6
7
class_names = train_ds_raw.class_names
num_classes = len(class_names)
print("类别:", class_names)

AUTOTUNE = tf.data.AUTOTUNE
train_ds = train_ds_raw.cache().shuffle(1000).prefetch(buffer_size=AUTOTUNE)
val_ds = val_ds_raw.cache().prefetch(buffer_size=AUTOTUNE)

查看不同标签图片的数量

1
2
3
4
5
6
7
import numpy as np
import matplotlib.pyplot as plt

counts = np.bincount([y for x, y in train_ds.unbatch()])
plt.bar(range(len(class_names)), counts)
plt.xticks(range(len(class_names)), class_names)
plt.show()

photo

数据增强

1
2
3
4
5
6
7
8
9
IMG_SIZE = (224, 224)

data_augmentation = tf.keras.Sequential([
    layers.RandomFlip("horizontal_and_vertical"),
    layers.RandomRotation(0.15), # 随机旋转
    layers.RandomZoom(0.2), # 随机缩放
    layers.RandomContrast(0.2), # 随机对比度
    layers.RandomBrightness(0.2) # 随机亮度
])

模型选择

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
# DenseNet121
base_model = tf.keras.applications.DenseNet121(
    include_top=False,
    weights='imagenet',
    input_shape=(224, 224, 3)
)
base_model.trainable = False

model = models.Sequential([
    data_augmentation,
    layers.Lambda(preprocess_input),
    base_model,
    layers.GlobalAveragePooling2D(),
    layers.Dropout(0.3),
    layers.Dense(128, activation='relu', kernel_regularizer=tf.keras.regularizers.l2(1e-4)),
    layers.Dense(4, activation='softmax')
])

迁移学习

第一阶段

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
model.compile(
    optimizer=tf.keras.optimizers.Adam(1e-3),
    loss='sparse_categorical_crossentropy',
    metrics=['accuracy']
)

print("第一阶段")

history1 = model.fit(
    train_ds,
    validation_data=val_ds,
    epochs=10,
    callbacks=[
        tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=3, restore_best_weights=True)
    ]
)

Epoch 1/10
2025-11-16 07:24:40.974143: I tensorflow/core/grappler/optimizers/custom_graph_optimizer_registry.cc:117] Plugin optimizer for device_type GPU is enabled.
264/264 ━━━━━━━━━━━━━━━━━━━━ 183s 645ms/step - accuracy: 0.6757 - loss: 0.8693 - val_accuracy: 0.8177 - val_loss: 0.4993
Epoch 2/10
264/264 ━━━━━━━━━━━━━━━━━━━━ 135s 511ms/step - accuracy: 0.7649 - loss: 0.6108 - val_accuracy: 0.8191 - val_loss: 0.5096
Epoch 3/10
264/264 ━━━━━━━━━━━━━━━━━━━━ 143s 540ms/step - accuracy: 0.7749 - loss: 0.5895 - val_accuracy: 0.8329 - val_loss: 0.4379
Epoch 4/10
264/264 ━━━━━━━━━━━━━━━━━━━━ 140s 531ms/step - accuracy: 0.7840 - loss: 0.5671 - val_accuracy: 0.7779 - val_loss: 0.5830
Epoch 5/10
264/264 ━━━━━━━━━━━━━━━━━━━━ 135s 511ms/step - accuracy: 0.7770 - loss: 0.5739 - val_accuracy: 0.8030 - val_loss: 0.4965
Epoch 6/10
264/264 ━━━━━━━━━━━━━━━━━━━━ 142s 539ms/step - accuracy: 0.7874 - loss: 0.5575 - val_accuracy: 0.8376 - val_loss: 0.4554

第二阶段

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
base_model.trainable = True

for layer in base_model.layers[:-80]:
    layer.trainable = False

print(f"第二阶段")
print(f"可训练参数: {sum([tf.size(v).numpy() for v in model.trainable_variables])}")

model.compile(
    optimizer=tf.keras.optimizers.Adam(1e-5),
    loss='sparse_categorical_crossentropy',
    metrics=['accuracy']
)

callbacks = [
    tf.keras.callbacks.EarlyStopping(
        monitor='val_loss', 
        patience=5, 
        restore_best_weights=True
    ),
    tf.keras.callbacks.ReduceLROnPlateau(
        monitor='val_loss', 
        factor=0.5, 
        patience=3,
        min_lr=1e-7
    )
]

history2 = model.fit(
    train_ds,
    validation_data=val_ds,
    epochs=40,
    callbacks=callbacks
)

…
Epoch 39/40
264/264 ━━━━━━━━━━━━━━━━━━━━ 273s 1s/step - accuracy: 0.9297 - loss: 0.2094 - val_accuracy: 0.9081 - val_loss: 0.2820 - learning_rate: 5.0000e-06
Epoch 40/40
264/264 ━━━━━━━━━━━━━━━━━━━━ 324s 1s/step - accuracy: 0.9315 - loss: 0.1987 - val_accuracy: 0.9058 - val_loss: 0.2970 - learning_rate: 5.0000e-06

验证

1
2
val_loss, val_acc = model.evaluate(val_ds_raw)
print(f"验证集准确率: {val_acc:.3f}")
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
import matplotlib.pyplot as plt

plt.figure(figsize=(12, 5))

# 准确率
plt.subplot(1, 2, 1)
plt.plot(history2.history["accuracy"], label="Train Accuracy")
plt.plot(history2.history["val_accuracy"], label="Val Accuracy")
plt.legend()
plt.title("Accuracy over Epochs")

# 损失
plt.subplot(1, 2, 2)
plt.plot(history2.history["loss"], label="Train Loss")
plt.plot(history2.history["val_loss"], label="Val Loss")
plt.legend()
plt.title("Loss over Epochs")

plt.show()

photo

保存模型

1
2
model.save("brain_tumor_model_3.keras")
print("模型已保存为 brain_tumor_model_3.keras")

加载模型

1
2
3
4
model = tf.keras.models.load_model(
    "brain_tumor_model_3.keras",
    custom_objects={"preprocess_input": preprocess_input}
)

测试

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
import numpy as np
import matplotlib.pyplot as plt

class_names = train_ds_raw.class_names

for images, labels in val_ds_raw.take(1):
    preds = model.predict(images)
    pred_labels = np.argmax(preds, axis=1)
    
    plt.figure(figsize=(12, 12))
    for i in range(9):
        plt.subplot(3, 3, i + 1)
        plt.imshow(images[i].numpy().astype("uint8"))
        true_label = class_names[labels[i]]
        pred_label = class_names[pred_labels[i]]
        color = "green" if true_label == pred_label else "red"
        plt.title(f"Pred: {pred_label}\nTrue: {true_label}", color=color)
        plt.axis("off")
    plt.show()

photo

分类报告

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.metrics import confusion_matrix, classification_report

y_true = []
y_pred = []

for images, labels in val_ds:
    preds = model.predict(images, verbose=0)
    preds_class = np.argmax(preds, axis=1)
    y_true.extend(labels.numpy())
    y_pred.extend(preds_class)

y_true = np.array(y_true)
y_pred = np.array(y_pred)

print("Classification Report:")
print(classification_report(y_true, y_pred, digits=4))

cm = confusion_matrix(y_true, y_pred)
cm_norm = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]

plt.figure(figsize=(8,6))
sns.heatmap(cm_norm, annot=True, fmt=".2f", cmap="Blues", cbar=True)
plt.title("Classification Report")
plt.show()

val_acc = np.mean(y_true == y_pred)
print(f"验证集准确率: {val_acc:.4f}")

photo

  • Title: brain tumor dataset
  • Author: 姜智浩
  • Created at : 2025-12-02 11:45:14
  • Updated at : 2025-12-02 10:36:41
  • Link: https://super-213.github.io/zhihaojiang.github.io/2025/12/02/20251202brain tumor dataset/
  • License: This work is licensed under CC BY-NC-SA 4.0.
On this page
brain tumor dataset
  1. 数据来源
  2. 导入库
  3. 检查是否启用了GPU
  4. 构造数据集
  5. 查看不同标签图片的数量
  6. 数据增强
  7. 模型选择
  8. 迁移学习
    1. 第一阶段
    2. 第二阶段
  9. 验证
  10. 保存模型
  11. 加载模型
  12. 测试
  13. 分类报告