CNN Semantic segmentation network (UNET) performing poorly when tested with images

Question

我正在构建一个 UNET 语义分割网络，为了检测眼睛图像中的虹膜，我们使用 DICE 损失函数，我们实现了超过 90% 的训练平均 IOU 和大约 87% 的验证数据，但验证损失和平均 IOU 在此期间显示出波动200 个时期，当我们测试我们的网络时，它显示出非常糟糕的结果，即使我们使用训练集中的图像，输出图像也与真实情况完全不同，它仍然表现不佳，关于我们可以做些什么来克服这个问题的任何建议.

Answer 1

Is the dataset public? You can find some good resources and guidance here: https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow2/Segmentation/UNet_Medical you will find also benchmark results. I think they are usinf ADAM optimizer with 0.0001 learning rate. More details in the training process section.

Answer 2

What does your implementation of the DSC Loss look like? And what is the activation function you're using? I've personally used:

def dice_coef(y_true, y_pred, smooth=1e-3):
    y_true_f = K.flatten(y_true)
    y_pred_f = K.flatten(y_pred)
    intersection = K.sum(y_true_f * y_pred_f)
    dice = (2. * intersection + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)
    return dice

def dice_coef_loss(y_true, y_pred):
    return 1 - dice_coef(y_true, y_pred)

as the loss function and metric in several U-Net segmentation architectures, and sigmoid as the activation function (assuming it's a binary classification problem). Also, if you're optimizing for IOU, you may want to change your loss to be an IOU loss - something along the lines of:

def iou(y_true, y_pred, smooth=1e-3):
    intersection = K.sum(K.abs(y_true * y_pred), axis=-1)
    union = K.sum((y_true,-1) + K.sum(y_pred,-1) - intersection
    return (intersection + smooth) / ( union + smooth)

def iou_loss(y_true, y_pred):
    return -iou_coef(y_true, y_pred)

Hope this helps!