Amshaker / unetr_plus_plus

[IEEE TMI-2024] UNETR++: Delving into Efficient and Accurate 3D Medical Image Segmentation
Apache License 2.0
345 stars 32 forks source link

Dataset of ACDC #19

Closed whduddhks closed 1 year ago

whduddhks commented 1 year ago

I downloaded the ACDC data set provided by README.md. The dataset.json in the ACDC dataset contains information on training and testing, and the data that needs to be tested is also written in training, so I wonder if the training code excludes it. Also, in nnFormer's dataset.json for ACDC dataset, the test data is not written in training, so I wonder why it was added.

In addition, if the test data was used for training, I wonder if the pre-trained model provided was learned in the same way.

Thank you.

Amshaker commented 1 year ago

Hi @whduddhks,

The training and testing data are written in json, and yes, the training code excludes it.

You can specify which samples are for testing in the ACDC trainer file: https://github.com/Amshaker/unetr_plus_plus/blob/024f24bd57425fc952eec6bf6ad02af6b1b2e017/unetr_pp/training/network_training/unetr_pp_trainer_acdc.py#L401

So the testing data is not used for training, it is just written in the JSON but the trainer file excludes it.

Please let me know if there is something unclear.

Regards, Abdelrahman.

whduddhks commented 1 year ago

Hi, @Amshaker

Thank you for your answer.

I understand that test data is not used in training.

And I have one more questions.

It is ok that the test data is the same as the validation data? Or Train data is split to train and validation when training? When I checked the code, test data is same as the validation data.

Thank you, whduddhks

Amshaker commented 1 year ago

Yes, it is ok.

However, you can divide the training set into train and another validation set to tune the hyper-parameters, and then use the testing set for evaluation. In this way, you have to update the trainer file to divide them accordingly.

Best regards, Abdelrahman.