使用 keras 在多类图像分割中加权样本

Question

我正在使用基于 Unet 的模型对生物医学图像执行图像分割。每个图像都是 224x224，我有四个类，包括背景类。每个掩码的大小为 (224x224x4)，因此我的生成器创建了一批大小为 (16x224x224x4) 的 numpy 数组。我将掩码的值重铸为 1 或 0，因此对于每个类，相关通道中都存在 1。图像也按 1/255 缩放。我使用骰子分数作为训练期间的性能指标，使用 1 骰子分数作为损失函数。我在训练期间的分数似乎高达 0.89，但我发现当我在测试集上预测时，我总是在预测背景类。我只在几百张图像上训练了 10 个 epochs（虽然我可以访问更多）这可能会影响模型，但我以为我' d 仍然得到其他类别的预测，所以我假设主要问题是类别不平衡。从在线查看 sample_weight 参数可能是答案，但我不确定我打算如何实现实际的重量部分？大概我需要使用图层将权重应用于模型中某个点的像素数组，但我不确定如何。任何帮助将非常感激？

class DataGenerator(keras.utils.Sequence):
     def __init__(self, imgIds, maskIds, imagePath, maskPath, batchSize=16, imageSize = (224, 224, 3), nClasses=2, shuffle=False):
       self.imgIds = imgIds
       self.maskIds = maskIds
       self.imagePath = imagePath
       self.maskPath = maskPath
       self.batchSize = batchSize
       self.imageSize = imageSize
       self.nClasses = nClasses
       self.shuffle = shuffle


     def __load__(self, imgName, maskName):

       img = cv2.imread(os.path.join(self.imagePath,imgName))
       img = cv2.resize(img, (self.imageSize[0], self.imageSize[1]))

       mask = cv2.imread(os.path.join(self.maskPath,maskName))
       mask = np.dstack((mask, np.zeros((4000, 4000))))

       mask[:,:,3][mask[:,:,0]==0]=255
       mask = mask.astype(np.bool)
       mask = img_as_bool(resize(mask, (self.imageSize[0], self.imageSize[1])))
       mask = mask.astype('uint8')

       img = img/255.0
       mask = mask

       return (img, mask)


    def __getitem__(self, index):

       if(index+1)*self.batchSize > len(self.imgIds):
          self.batchSize = len(self.imgIds) - index*self.batchSize

       batchImgs = self.imgIds[self.batchSize*index:self.batchSize*(index+1)]
       batchMasks = self.maskIds[self.batchSize*index:self.batchSize*(index+1)]

       batchfiles = [self.__load__(imgFile, maskFile) for imgFile, maskFile in 
       zip(batchImgs, batchMasks)]

       images, masks = zip(*batchfiles)

       return np.array(list(images)), np.array(list(masks))


   def __len__(self):
       return int(np.ceil(len(self.imgIds)/self.batchSize))


class Unet():
   def __init__(self, imgSize):
       self.imgSize = imgSize


   def convBlocks(self, x, filters, kernelSize=(3,3), padding='same', strides=1):

       x = keras.layers.BatchNormalization()(x)
       x = keras.layers.Activation('relu')(x)
       x = keras.layers.Conv2D(filters, kernelSize, padding=padding, strides=strides)(x)

       return x


   def identity(self, x, xInput, f, padding='same', strides=1):

      skip = keras.layers.Conv2D(f, kernel_size=(1, 1), padding=padding, strides=strides)(xInput)
      skip = keras.layers.BatchNormalization()(skip)
      output = keras.layers.Add()([skip, x])

      return output


    def residualBlock(self, xIn, f, stride):

      res = self.convBlocks(xIn, f, strides=stride)
      res = self.convBlocks(res, f, strides=1)
      output = self.identity(res, xIn, f, strides=stride)

      return output


    def upSampling(self, x, xInput):

      x = keras.layers.UpSampling2D((2,2))(x)
      x = keras.layers.Concatenate()([x, xInput])

      return x


    def encoder(self, x, filters, kernelSize=(3,3), padding='same', strides=1):

      e1 = keras.layers.Conv2D(filters[0], kernelSize, padding=padding, strides=strides)(x)
      e1 = self.convBlocks(e1, filters[0])

      shortcut = keras.layers.Conv2D(filters[0], kernel_size=(1, 1), padding=padding, strides=strides)(x)
      shortcut = keras.layers.BatchNormalization()(shortcut)
      e1Output = keras.layers.Add()([e1, shortcut])

      e2 = self.residualBlock(e1Output, filters[1], stride=2)
      e3 = self.residualBlock(e2, filters[2], stride=2)
      e4 = self.residualBlock(e3, filters[3], stride=2)
      e5 = self.residualBlock(e4, filters[4], stride=2)

      return e1Output, e2, e3, e4, e5


  def bridge(self, x, filters):

      b1 = self.convBlocks(x, filters, strides=1)
      b2 = self.convBlocks(b1, filters, strides=1)

      return b2


  def decoder(self, b2, e1, e2, e3, e4, filters, kernelSize=(3,3), padding='same', strides=1):

      x = self.upSampling(b2, e4)
      d1 = self.convBlocks(x, filters[4])
      d1 = self.convBlocks(d1, filters[4])
      d1 = self.identity(d1, x, filters[4])

      x = self.upSampling(d1, e3)
      d2 = self.convBlocks(x, filters[3])
      d2 = self.convBlocks(d2, filters[3])
      d2 = self.identity(d2, x, filters[3])

      x = self.upSampling(d2, e2)
      d3 = self.convBlocks(x, filters[2])
      d3 = self.convBlocks(d3, filters[2])
      d3 = self.identity(d3, x, filters[2])

      x = self.upSampling(d3, e1)
      d4 = self.convBlocks(x, filters[1])
      d4 = self.convBlocks(d4, filters[1])
      d4 = self.identity(d4, x, filters[1])

      return d4 


  def ResUnet(self, filters = [16, 32, 64, 128, 256]):

      inputs = keras.layers.Input((224, 224, 3))

      e1, e2, e3, e4, e5 = self.encoder(inputs, filters)
      b2 = self.bridge(e5, filters[4])
      d4 = self.decoder(b2, e1, e2, e3, e4, filters)

      x = keras.layers.Conv2D(4, (1, 1), padding='same', activation='softmax')(d4)
      model = keras.models.Model(inputs, x)

      return model


imagePath = 'output/t2'
maskPath = 'output/t1'

imgIds = glob.glob(os.path.join(imagePath, '*'))
maskIds = glob.glob(os.path.join(maskPath, '*'))

imgIds = [os.path.basename(f) for f in imgIds]
maskIds = [os.path.basename(f) for f in maskIds]

trainImgIds = imgIds[:300]
trainMaskIds = maskIds[:300]
validImgIds = imgIds[300:350]
validMaskIds = maskIds[300:350]

trainGenerator = DataGenerator(trainImgIds, trainMaskIds, imagePath, maskPath, **params)
validGenerator = DataGenerator(validImgIds, validMaskIds, imagePath, maskPath)

trainSteps = len(trainImgIds)//trainGenerator.batchSize
validSteps = len(validImgIds)//validGenerator.batchSize

unet = Unet(224)
model = unet.ResUnet()
model.summary()

adam = keras.optimizers.Adam()
model.compile(optimizer=adam, loss=dice_coef_loss, metrics=[dice_coef])

hist = model.fit_generator(trainGenerator, validation_data=validGenerator, 
steps_per_epoch=trainSteps, validation_steps=validSteps, 
                verbose=1, epochs=6)

Answer 1

为了跟进这一点，我使用sample_weight让它工作。如果您知道自己必须做什么，那就太好了。不幸的是，文档对此并不清楚，大概是因为该功能最初是为时间序列数据添加的。

• 当您指定模型时，您需要在损失函数之前将 2D 图像大小的输出重塑为向量。
• 编译模型时使用sample_weight_mode =“temporal”。这将允许您传入一个权重矩阵进行训练，其中每一行代表单个样本的权重向量。

我希望这有帮助。