search

taigw / brats17

Brain tumor segmentation for MICCAI 2017 BraTS challenge
BSD 3-Clause "New" or "Revised" License
321 stars 130 forks source link

the validation data segmentation result #25

Open Qingyuncookie opened 5 years ago

Qingyuncookie commented 5 years ago

I have tested your given models under the model17 directory on the validation data by submitting the segmentation results to the CBICA Image Processing Portal. However, I didn't get the same result you presented in the paper. The dice values and Hausdorff values are (0.7572, 0.8989, 0.8350) and ( 3.7947, 5.7439, 7.3037) respectively, In your paper, the values are (0.7859, 0.9050, 0.8378) and (3.2821, 3.8901, 6.4790) respectively. Would you like tell me if you apply any other methods such as model ensemble to get the result in the paper?

1160914483 commented 5 years ago

I used the original configuration file and the corresponding parameters, but the score on the verification set was (0.75315 0.90392 0.80955). I'd like to ask the same question.

taigw commented 5 years ago

The pre-trained models released here are not exactly the ones I used in the paper. To release the repository, I re-organized the code to make it clearer to understand, then I re-trained the model. However, the re-training was implemented on another GPU with 6GB memory so I reduced the batch size. This caused a performance decrease.

Qingyuncookie commented 5 years ago

@taigw Thank you for your reply! When I trained the model, I set the batch_size to 3 and keep all other options unchanged in the configure files. And the result I got is (0.7376, 0.8997, 0.7912) after 20000 iterations, it's still lower than the result tested by the pre-trained model you gave, I don't know why the model I trained can't reproduce the result of your‘s, is there anything wrong with what I didn't pay attention to?