UNET won't even overfit on 8 examples from BraTS dataset

I have been trying to train some image segmentation models on the BraTS23 dataset, unfortunately I couldn’t get good results, so I started to doubt my code or the way I calculate the metrics.
then I decided to train a UNET on the 8 first examples (which translates to about 1000 4channels 2d inputs) from this dataset, my goal is to overfit this model, so at least I can say the model can learn, but even with this it won’t overfit, I used the mean IOU as metric and the best I can get is 33 % after 100 epoch.

I have prepared a colab notebook if you would like to take a look at my code: