How to train a classifier for a situation that has imbalanced data using the support provided by MONAI?

[DSRajesh]

Hello

We are building a 3D volume classifier for classifying an object image into one of three classes. We expect the classifier to output 3 categorical outputs each one corresponding to one of the 3 classes. The dataset we are using is imbalanced such that object images (3D) corresponding to the first class are plenty in number when compared to the number of the other two classes i.e the numbers of the images of 3 classes is approximately 900, 57, 38 respectively. We have tried with DenseNet121 network and some simple augmentation techniques provided on MONAI website, like random rotation and scaling, but not with much success. If you could please suggest some more utilities of MONAI to handle data imbalance (like stratified sampling, positive-negative sampling (hard_negative_sampler)) such a situation it will be helpful. Thanking you

Rajesh

[KumoLiu]

Hi @DSRajesh, thanks for your interest here.
You can try these two transforms according to your situation.

MONAI/monai/transforms/croppad/array.py

Line 1049 in 4743945

class RandCropByPosNegLabel(Randomizable, TraceableTransform, LazyTransform, MultiSampleTrait):

MONAI/monai/transforms/croppad/array.py

Line 1228 in 4743945

class RandCropByLabelClasses(Randomizable, TraceableTransform, LazyTransform, MultiSampleTrait):

Hope it help, thanks!