clDice - a Novel Topology-Preserving Loss Function for Tubular Structure Segmentation

First of all - thanks for a great paper about clDice, it's really interesting approach.

I wanted to test the idea on 3D dataset.

I have a synthetic 3D shape on which I just run soft_skeletonize and I assume it should leave the same single connected component. unfortunately it doesn't. See the following summary showing the input image on the left, and two iterations of soft_skel - for both tensorflow and pytorch.

I've created a reproducible Colab notebook for the case: https://gist.github.com/kretes/84f6025e7e1ded19591a54b62abcc539

https://github.com/jocpae/clDice/blob/e9a5f8efc376eb1961b276e698a38556224b7d60/cldice_metric/cldice.py#L32

the definition of precision and sensitivity seems not consist with cldice loss and common definition, although it will not affect the final cldice score tprec = cl_score(v_l,skeletonize_3d(v_p)) tsens = cl_score(v_p,skeletonize_3d(v_l))

Hi, I really appreciate your work. Can you please also share the implementation of topology-based evaluation metrics that you used in your paper? That will be very helpful. will look forward to your response. Thanks.

During training, should I convert the output from the Unet to a binary mask prior feeding into soft_dice_cldice? I used raw logits, but the training result is really bad. However when I tried to convert to a binary mask, I lost the grad on the tensor.

Thank you very much for your article. What I want to ask is whether cldice is used as a module? If I have multiple tasks, how should I combine multiple loss functions for training?

Hi, thanks for the great package.

I would like to understand why you perform soft erosion using a sequence of 3 min_pooling with separable filters. Can't we instead just use one big min_pool (using -F.max_pool3d(-img, kernel_size=(3,3,3),stride=1,padding=1) ) ?

When i use the cldice to train a vessel segmentation network, i find that the cldice keeps 0 all the time.

class soft_cldice(nn.Module): def init(self, iter_=3, smooth = 1.): super(soft_cldice, self).init() self.iter = iter_ self.smooth = smooth

def forward(self, y_pred, y_true):
    y_true = y_true.contiguous().unsqueeze(1).to(float) # to get the true label
    y_pred = (y_pred > 0.5).contiguous().to(float).requires_grad_() #to get the pred mask
    skel_pred = soft_skel(y_pred, self.iter)
    skel_true = soft_skel(y_true, self.iter)
    tprec = (torch.sum(torch.multiply(skel_pred, y_true)[:,1:,...])+self.smooth)/(torch.sum(skel_pred[:,1:,...])+self.smooth)    
    tsens = (torch.sum(torch.multiply(skel_true, y_pred)[:,1:,...])+self.smooth)/(torch.sum(skel_true[:,1:,...])+self.smooth)    
    cl_dice = 1.- 2.0*(tprec*tsens)/(tprec+tsens)
    return cl_dice

Can you provide an example of cldice as a loss function to train the network? I used cldice as a loss function for training, but the result is that I can't train it. The training loss curve is also very strange, and even negative values, and the shock is large, and it cannot be converged.

Thanks for a great paper about clDice. I am a newbie who just started deep learning. I want to add clDice as the loss function of the network. How should I implement it? Besides, Can it be combined with other loss functions？

Hello, how did you get the boundary mask of CREMI? It only provides the neuron_ids. I can extract the boundaries by applying skimage.segmentation.find_boundarys on neuron_ids. But I got boundaries which seem thinner than what you have shown in paper. Did you add something such as a Gaussian blur? Thanks!

Hello, when I use soft_skeleton, I found only large vessels left no matter how I set the iteration, please tell me what should I do? I'm a little confused about the theory. Thanks!

Hello nice work!

I was wondering if this is the right implementation of the clDice metrics. Your paper indicated high performance of this metric. However, finding the clDice score of a vessel mask between itself was resulting 91%.

I used your implementation by the way. Other metrics were reporting 100%. Have any idea what the bug could be?