DenseNet121 overfitting on I-SPY 1 3D data classification

[jamesben6688]

I am using the DenseNet 121 (3D) to do the 3D MRI volume classification (pCR prediction) on the I-SPY 1 dataset (https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=20643859), by using the T2 post-contrast images. However the validation result is terrible (less than 70%). Moreover, the model will mostly output 0, I already used the weighted crossentropy.

Have anyone done experiments on such data, are there anything special I should take care of?

[KumoLiu]

Hi @jamesben6688, if overfitting, perhaps you could try using a simpler network such as vgg. And you can also follow some data preparation introduced in https://github.com/MIC-DKFZ/nnUNet/tree/master such as how to do normalization.
Hope it can help you, thanks!