I have a highly imbalanced data-distributon(3-4% of total voxels represent the object of interest) for my problem. I would like to know how to enable Niftynet's balanced window sampler here? Do I need to modify the get_subimage_batch() method?
Further, would you please suggest other features of your code which can be used to address the class-imbalance problem. I am using ROI patches for training with improved results and planning to use the tversky loss function in future.
Hi,
I have a highly imbalanced data-distributon(3-4% of total voxels represent the object of interest) for my problem. I would like to know how to enable Niftynet's balanced window sampler here? Do I need to modify the
get_subimage_batch()
method?Further, would you please suggest other features of your code which can be used to address the class-imbalance problem. I am using ROI patches for training with improved results and planning to use the tversky loss function in future.