Results after training

[AMagd]

First of all, thanks for publicly sharing your paper and source-code.

I have used train.py to train a model on BraTS2020 dataset, but the results I got mismatches the results reported in the paper, do you have any idea why is that?

I got the following Dice scores (after training for 300 epochs): wt is 0.8935, tc is 0.7762, et is 0.7637, mean_dice is 0.8111

[920232796]

Can you show the tensorboard training figures? And what the validation dice scores? The mean validation dice should be about 0.87.

[AMagd]

Thanks for your reply.

1. Here are the tensorboard plots:

2. The average test dice score over the 75 sample (by running test.py) is: wt = 0.9126328895901329 tc = 0.8598519600752457 et = 0.7791602345663959 mean_dice = 0.8505483614105913

The mean dice for the test dataset (0.85) is much better than the (0.81) I got during training

[920232796]

Your training result seems like a bit strange.

[920232796]

My tensorboard result in training is :

[AMagd]

In the image in your first reply, I've noticed that the folder name you have is different than mine "diffusion_seg_all_loss_e300_norm_fix...", so I am assuming that this means that the result you are showing are for a different (maybe a bit enhanced) model, is that correct? or are you getting these results from the exact same code published here?

[920232796]

No, the code is same. You maybe need to check the data? Or you can try to train on the BTCV dataset. I can also provide some information for you.

[AMagd]

Hmmm strange, I will double check once again then. Thanks a lot :)