results about focal-tversky-unet

[nabsabraham]

Hi jie, I appreciate your interest in our work. I would like to ask that you limit your issues to one thread because you have re-opened issues that are not related to code reproducibility is in #11 and #10.
With regards to BUS code reproducibility, I have resampled the images to 128x128 and I used opencv to do this resizing. This could be a source of error. The other thing is that I used early stopping to pick the model with the best performance because of training instability when close to convergence. A last note is that I ran this script 5 times to get the DSC +/- some variance - in doing so, the train and test was split randomly so this is another area for randomness. I see that your result of 0.748 is still quite far from my result so I will have to look for the pre-processing code because that is the only thing I have not shown in this repo.

Can you try to reproduce the ISIC results? I have recently moved to pytorch and have reproduced ISIC without any struggles.

[jiejiangd]

Thank you for your reply!First of all, you have processed the data to 128 * 128 in the bus code, but there is no relevant processing for the ISIC code, so I think it is easier to reproduce the bus code than isic's. Secondly, I think the cross validation should not be too different, but I will review the code well according to your ideas.

In addition, is it convenient for you to share your Pytorch code? Maybe I'm a little more familiar with Pytorch. It doesn't matter if it's inconvenient. Thank you.

[ykeivn]

Hi, I try to use my data with this code, but why I get the dsc>1. Like this:
16/112 [===>..........................] - ETA: 11s - loss: -0.1059 - dsc: 1.1925 - tp: 1.0000 - tn: 3.8405e-06
32/112 [=======>......................] - ETA: 9s - loss: -0.1233 - dsc: 1.2266 - tp: 1.0000 - tn: 3.8427e-06
48/112 [===========>..................] - ETA: 7s - loss: -0.1497 - dsc: 1.2816 - tp: 1.0000 - tn: 3.8468e-06
64/112 [================>.............] - ETA: 5s - loss: -0.1684 - dsc: 1.3214 - tp: 1.0000 - tn: 3.8500e-06
80/112 [====================>.........] - ETA: 3s - loss: -0.1762 - dsc: 1.3380 - tp: 1.0000 - tn: 3.8510e-06
96/112 [========================>.....] - ETA: 1s - loss: -0.1816 - dsc: 1.3492 - tp: 1.0000 - tn: 3.8519e-06
112/112 [==============================] - 14s 125ms/step - loss: -0.1881 - dsc: 1.3637 - tp: 1.0000 - tn: 3.8533e-06 - val_loss: -0.2142 - val_dsc: 1.4020 - val_tp: 1.0000 - val_tn: 9.6319e-06
Can you give me some advice to solve the problem? Think you a lot in advance.

[nabsabraham]

Hi Kevin, can you check the range of your predictions and the range of your ground truth? Both masks should be in range [0-1]. You can just print the max and check what it is. Also, try adjusting the alpha and gamma parameters. Maybe your dataset is easier to segment and so the network is becoming super biased to prediction TPs and FPs.

[luistelmocosta]

Hey @nabsabraham, any chance that you can share the pytorch version? I really would like to try it

[nabsabraham]

Hi @luistelmocosta, thanks for your interest! Is it the model you were looking for or just the loss function? If it is the latter, I just wrote this up quickly but I think it should work:

def ftl(pred, gt, alpha=0.7, gamma=0.75):
    pflat = pred.contiguous().view(-1)
    gtflat = gt.contiguous().view(-1)
    
    intersection = (pflat * gtflat).sum()
    TP = intersection
    FP = (pflat * (1-gtflat)).sum()
    FN = ((1-pflat) * gtflat).sum()
    
    return 1 - (((TP + smooth)/(TP + alpha*FN + (1-alpha)*FP + smooth))**gamma).mean()

If it's the model, I believe you can get started with oktay's pytorch version of attention networks. I will have to rewrite it because I never hung on to it but its essentially a few extra layers added on to oktay's model.

[luistelmocosta]

Thank you for the quick reply.
I was following this one https://github.com/LeeJunHyun/Image_Segmentation and yea, I noticed that your Attn Unet have an extra layer (4 skip connections), should be easy to adapt.

I am currently using this version:

class TverskyLoss(nn.Module):
    __name__ = 'tversky_loss'
    def __init__(self, activation = None, apply_nonlin=None, batch_dice=False, do_bg=True, smooth=1.,
                 square=False):
        
        
        """
        paper: https://arxiv.org/pdf/1706.05721.pdf
        """
        super(TverskyLoss, self).__init__()
        
        

        self.square = square
        self.do_bg = do_bg
        self.batch_dice = batch_dice
        self.apply_nonlin = apply_nonlin
        self.smooth = smooth
        self.alpha = tversky_alpha
        self.beta = tversky_beta
        self.activation = Activation(activation)
        

    def forward(self, y_pr, y_gt, loss_mask=None):
        shp_x = y_pr.shape

        if self.batch_dice:
            axes = [0] + list(range(2, len(shp_x)))
        else:
            axes = list(range(2, len(shp_x)))

        if self.apply_nonlin is not None:
            y_pr = self.apply_nonlin(y_pr)
        
        y_pr = self.activation(y_pr)
        #tp, fp, fn = get_tp_fp_fn(x, y, axes, loss_mask, self.square)
        tp = torch.sum(y_gt * y_pr)
        fp = torch.sum(y_pr) - tp
        fn = torch.sum(y_gt) - tp 


        tversky = (tp + self.smooth) / (tp + self.alpha*fp + self.beta*fn + self.smooth)

        if not self.do_bg:
            if self.batch_dice:
                tversky = tversky[1:]
            else:
                tversky = tversky[:, 1:]
        tversky = tversky.mean()

        return -tversky

## Tverky Focal Loss

class FocalTversky_loss(nn.Module):
    __name__ = 'tversky_focal_loss'
    def __init__(self, tversky_kwargs, gamma=float(4/3)):
        super(FocalTversky_loss, self).__init__()
        self.gamma = gamma
        self.tversky = TverskyLoss()
        #print("t loss" , self.tversky)
        

    def forward(self, net_output, target):
        #print(self.tversky(net_output, target))
        tversky_loss = 1 + self.tversky(net_output, target) # = 1-tversky(net_output, target)
        focal_tversky = torch.pow(tversky_loss, self.gamma)
        return focal_tversky

Does this look good or should I had contiguous to my } y_gt and y_pr?

Thank you