基于胸片的COVID-19检测CNN模型过拟合问题求助

我正在构建一个基于CNN的胸片COVID-19检测模型，使用包含约27000张图像的数据集，仅选用其中的COVID和NORMAL两类。参考Keras图像分类教程调整后得到代码如下：

batch_size = 16
img_height = 160
img_width = 160
img_size = (img_height, img_width)

seed_train_validation = 1
shuffle_value = True
validation_split = 0.3

train_ds = tf.keras.utils.image_dataset_from_directory(
    data_dir,
    image_size = img_size,
    validation_split = validation_split,
    subset = "training",
    seed = seed_train_validation,
    color_mode = "grayscale",
    shuffle = shuffle_value
)

val_ds = tf.keras.utils.image_dataset_from_directory(
    data_dir,
    image_size = img_size,
    validation_split = validation_split,
    subset = "validation",
    seed = seed_train_validation,
    color_mode = "grayscale",
    shuffle = shuffle_value
)

val_batches = tf.data.experimental.cardinality(val_ds)
test_ds = val_ds.take((2*val_batches) // 3)
val_ds = val_ds.skip((2*val_batches) // 3)

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

resize_and_rescale = tf.keras.Sequential([
  layers.Resizing(img_height, img_width),
  layers.Rescaling(1./255)
])
data_augmentation = tf.keras.Sequential([
  layers.RandomFlip("horizontal_and_vertical"),
  layers.RandomRotation(0.2),
  layers.RandomZoom(0.1)
])

num_classes = len(class_names)

model_1 = Sequential([
    resize_and_rescale,
    data_augmentation,
    layers.Conv2D(16, 3, padding='same', activation='relu'),
    layers.MaxPooling2D(),
    layers.Conv2D(32, 3, padding='same', activation='relu'),
    layers.MaxPooling2D(),
    layers.Conv2D(64, 3, padding='same', activation='relu'),
    layers.MaxPooling2D(),
    layers.Dropout(0.2),
    layers.Flatten(),
    layers.Dense(128, activation='relu'),
    layers.Dense(num_classes)
])
model_1.compile(optimizer="adam",
                loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
                metrics=['accuracy'])
epochs = 75
history = model_1.fit(
    train_ds,
    validation_data = val_ds,
    epochs = epochs
)

训练轮数较少（如10轮）时，准确率和损失曲线表现正常；但训练75轮时曲线出现异常，已加入数据增强和Dropout层但效果不佳，怀疑是过拟合问题，寻求解决建议。

• 提前停止训练（Early Stopping）
  直接避免模型过度训练，当验证集性能连续多轮无提升时自动停止，并恢复到最优权重：

  early_stopping = tf.keras.callbacks.EarlyStopping(
      monitor='val_loss',
      patience=10,
      restore_best_weights=True
  )
  # 训练时加入回调
  history = model_1.fit(
      train_ds,
      validation_data=val_ds,
      epochs=75,
      callbacks=[early_stopping]
  )
  

• 强化正则化
  现有正则化力度不足，可从两方面优化：

  • 提升Dropout比例至0.3-0.5
  • 给卷积层和全连接层添加L2正则化，限制权重规模：

    from tensorflow.keras import regularizers
    
    model_1 = Sequential([
        resize_and_rescale,
        data_augmentation,
        layers.Conv2D(16, 3, padding='same', activation='relu', kernel_regularizer=regularizers.l2(0.001)),
        layers.MaxPooling2D(),
        layers.Conv2D(32, 3, padding='same', activation='relu', kernel_regularizer=regularizers.l2(0.001)),
        layers.MaxPooling2D(),
        layers.Conv2D(64, 3, padding='same', activation='relu', kernel_regularizer=regularizers.l2(0.001)),
        layers.MaxPooling2D(),
        layers.Dropout(0.3),
        layers.Flatten(),
        layers.Dense(128, activation='relu', kernel_regularizer=regularizers.l2(0.001)),
        layers.Dense(num_classes)
    ])
    

• 优化数据增强策略
  胸片垂直翻转无医学意义，反而引入噪声，同时可增加更多有效扰动：

  data_augmentation = tf.keras.Sequential([
    layers.RandomFlip("horizontal"),
    layers.RandomRotation(0.15),
    layers.RandomZoom(0.15),
    layers.RandomContrast(0.1),
    layers.RandomTranslation(height_factor=0.1, width_factor=0.1)
  ])
  

• 精简模型结构
  降低模型容量减少过拟合风险：

  • 把全连接层的128个神经元缩减至64
  • 在卷积层后添加BatchNormalization层，稳定训练同时抑制过拟合：

    model_1 = Sequential([
        resize_and_rescale,
        data_augmentation,
        layers.Conv2D(16, 3, padding='same', activation='relu'),
        layers.BatchNormalization(),
        layers.MaxPooling2D(),
        layers.Conv2D(32, 3, padding='same', activation='relu'),
        layers.BatchNormalization(),
        layers.MaxPooling2D(),
        layers.Conv2D(64, 3, padding='same', activation='relu'),
        layers.BatchNormalization(),
        layers.MaxPooling2D(),
        layers.Dropout(0.3),
        layers.Flatten(),
        layers.Dense(64, activation='relu'),
        layers.Dense(num_classes)
    ])
    

• 调整学习率策略
  降低初始学习率或添加学习率衰减，避免训练后期震荡：

  lr_scheduler = tf.keras.optimizers.schedules.ExponentialDecay(
      initial_learning_rate=0.001,
      decay_steps=1000,
      decay_rate=0.96,
      staircase=True
  )
  optimizer = tf.keras.optimizers.Adam(learning_rate=lr_scheduler)
  model_1.compile(optimizer=optimizer,
                  loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
                  metrics=['accuracy'])
  

内容的提问来源于stack exchange，提问作者Vinicius Cavalcante