Unet: with augmentation and dropout 0.5 loss_per_im is worse no augmentation, dropout 0.3 mean subtraction and std deviation make a huge difference for deepmedic