### U-Net 实现完整流程 以下是基于 Python 的 U-Net 完整实现方案，分为三个主要部分：数据预处理、模型训练和模型测试。 --- #### 一、数据预处理 此阶段的目标是对 NIFTI 格式的图像和掩码进行读取、尺寸调整以及数据集划分。代码示例如下： ```python import os import numpy as np from skimage.transform import resize import nibabel as nib from sklearn.model_selection import train_test_split def preprocess_nifti(image_path, mask_path, target_shape=(256, 256)): """ 预处理NIFTI格式的图像和掩码。 :param image_path: 图像路径 :param mask_path: 掩码路径 :param target_shape: 调整后的目标形状，默认为 (256, 256) :return: 处理后的图像数组和掩码数组 """ img = nib.load(image_path).get_fdata() msk = nib.load(mask_path).get_fdata() # 归一化图像像素值到 [0, 1] img_normalized = (img - np.min(img)) / (np.max(img) - np.min(img)) # 将图像和掩码调整为目标尺寸 resized_img = resize(img_normalized, (*target_shape, img.shape[-1]), mode='constant', preserve_range=True) resized_msk = resize(msk, (*target_shape, msk.shape[-1]), mode='constant', preserve_range=True) return resized_img, resized_msk def split_dataset(images_dir, masks_dir, test_size=0.2): """ 划分数据集为训练集和验证集。 :param images_dir: 图像目录 :param masks_dir: 掩码目录 :param test_size: 测试集比例，默认为 0.2 :return: 训练集和测试集的图像与掩码列表 """ image_files = sorted(os.listdir(images_dir)) mask_files = sorted(os.listdir(masks_dir)) X_train, X_val, y_train, y_val = train_test_split( image_files, mask_files, test_size=test_size, random_state=42 ) return X_train, X_val, y_train, y_val ``` 以上代码实现了对 NIFTI 文件的加载、归一化、尺寸调整以及数据集划分功能[^1]。 --- #### 二、模型训练 在此阶段构建 U-Net 模型并完成训练过程。以下是一个简单的 U-Net 架构定义及训练脚本： ```python import tensorflow as tf from tensorflow.keras.layers import Input, Conv2D, MaxPooling2D, Dropout, concatenate, UpSampling2D from tensorflow.keras.models import Model def unet_model(input_size=(256, 256, 1), num_classes=1): inputs = Input(input_size) # 下采样层 conv1 = Conv2D(64, 3, activation='relu', padding='same')(inputs) conv1 = Conv2D(64, 3, activation='relu', padding='same')(conv1) pool1 = MaxPooling2D(pool_size=(2, 2))(conv1) conv2 = Conv2D(128, 3, activation='relu', padding='same')(pool1) conv2 = Conv2D(128, 3, activation='relu', padding='same')(conv2) pool2 = MaxPooling2D(pool_size=(2, 2))(conv2) # 中间桥接层 conv3 = Conv2D(256, 3, activation='relu', padding='same')(pool2) conv3 = Conv2D(256, 3, activation='relu', padding='same')(conv3) # 上采样层 up4 = UpSampling2D(size=(2, 2))(conv3) merge4 = concatenate([conv2, up4], axis=-1) conv4 = Conv2D(128, 3, activation='relu', padding='same')(merge4) conv4 = Conv2D(128, 3, activation='relu', padding='same')(conv4) up5 = UpSampling2D(size=(2, 2))(conv4) merge5 = concatenate([conv1, up5], axis=-1) conv5 = Conv2D(64, 3, activation='relu', padding='same')(merge5) conv5 = Conv2D(64, 3, activation='relu', padding='same')(conv5) outputs = Conv2D(num_classes, 1, activation='sigmoid')(conv5) model = Model(inputs=[inputs], outputs=[outputs]) model.compile(optimizer=tf.keras.optimizers.Adam(), loss='binary_crossentropy', metrics=['accuracy']) return model # 加载数据并训练模型 X_train, X_val, y_train, y_val = ... # 使用前面的函数获取数据 model = unet_model() history = model.fit(X_train, y_train, validation_data=(X_val, y_val), epochs=10, batch_size=8) ``` 该代码片段展示了如何创建一个标准的 U-Net 结构，并配置损失函数和优化器用于训练[^2]。 --- #### 三、模型测试 最后一步是在测试集上评估模型性能。以下是预测和可视化结果的代码示例： ```python import matplotlib.pyplot as plt def predict_and_visualize(model, test_images, test_masks): predictions = model.predict(test_images) fig, ax = plt.subplots(len(test_images), 3, figsize=(10, len(test_images)*5)) for i in range(len(test_images)): ax[i][0].imshow(np.squeeze(test_images[i]), cmap="gray") # 原始图像 ax[i][1].imshow(np.squeeze(predictions[i]), cmap="gray") # 预测结果 ax[i][2].imshow(np.squeeze(test_masks[i]), cmap="gray") # 真实掩码 plt.show() test_images, test_masks = ... # 准备测试数据 predict_and_visualize(model, test_images, test_masks) ``` 这段代码能够帮助直观比较原始图像、真实掩码和模型预测的结果[^3]。 --- ### 总结 上述代码涵盖了从数据预处理到模型训练再到最终测试的整个工作流。每一步都经过精心设计以确保高效性和准确性。