Understanding how to label/target tensors for 3D volumes

[moreshud]

Thanks for the explanation.

@ptrblck

usually the dice loss needs predictions, i.e. you would apply a threshold on the sigmoid of your output. However, this depends on your implementation of the dice loss. The shapes should be the same for your output and target.

Do you mean I need to threshold the prediction (i.e the probabilities after the activation function (sigmod/softmax)) so as to be one-hot encoded like the target?

Sample code:

def dice_single_channel(probability, truth, activation, eps = 1e-7, squared_pred = False):
    if activation:
        probability = nn.Softmax2d()(probability)

    p = (probability >= .5).float()
    t = (truth).float()

    intersection = (p * t).sum()

    if squared_pred:
        p = p**2
        t = t**2
 
    return 1-(2.0 * intersection + eps)/ (p.sum() + t.sum() + eps)

batchsize = 2
channel_num = 5
probability = torch.rand(batchsize,channel_num,10,10)

truth = torch.randint(low = 0, high=2, size=(batchsize,channel_num,10,10))
print(probability.sum(), truth.sum())

dice = dice_single_channel(probability, truth, activation=True) 
print(dice)

Outputs:

tensor(0.9302)

[ptrblck]

I’m not sure which implementation I was referring to, but would recommend to take a look at this feature request or this implementation, which don’t use a threshold.