The default 5-fold cross-validation split

[tarash66]

Dear,

I would be so grateful if you could respond to my question. As I figured out, the default 5-fold cross-validation split works randomly on the dataset. Is it so? If yes, would you please inform me whether the split for the training of Task082_BraTS2020 was also random?

Regards,
Tarash

[FabianIsensee]

Splits are always random. How else would we get reliable results?

[tarash66]

Dear Fabian.

I am so grateful for your response. I thought maybe you tried to keep a pattern like dividing the first 20 percent of the dataset for test and the rest for training in the first fold, then the second 20 percent of the dataset for testing in the 2nd fold, and so on and so forth until the 5th fold. Unfortunately, I couldn't find any record of images in each split. I would appreciate it if you could inform me where I can see it.

Regards,

[FabianIsensee]

See splits_final.json in your preprocessed data folder

[tarash66]

I mean your "splits_final.json" for BraTS2020. I am intersted to know what images from BraTs2020 dataset were used for each provided model (out of 5). I appericate your response.

Regards,

[FabianIsensee]

Ah I see.
splits_final.zip

Take this.
I had to zip it because GitHub doesn't allow pkl files...

[FabianIsensee]

All models are trained as 5 fold cross-validation with the splits from this file

[tarash66]

thanks dear Fabian. All the bests.