Try the SMOTE oversampling technique where the minority class is oversampled to increase the number of instances of that class:

https://en.wikipedia.org/wiki/Oversampling_and_undersampling_in_data_analysis

Dear Gunjan Naik ,

I think this previous discussion can help you:

https://www.quora.com/Can-deep-learning-handle-imbalanced-data#

Hi,

The advanced or old augmentation methods can be used to balance your data.

https://medium.com/nanonets/how-to-use-deep-learning-when-you-have-limited-data-part-2-data-augmentation-c26971dc8ced

https://www.researchgate.net/post/Machine_learning_if_proportion_of_number_of_cases_in_different_class_in_training_set_matters

https://arxiv.org/pdf/1801.05365

https://machinelearningmastery.com/tactics-to-combat-imbalanced-classes-in-your-machine-learning-dataset/

Actually there are a number of workarounds. 1st you may handle the issue with data augmentation of any kind. Or you may balance the dataset simply shrinking the negative class (if the dataset size allows to do so). 2nd you may tune the outcome. Remember that the network is just a classifier that outputs the positive class probability. Its your choice actually which probability threshold to use when you turn it to binary outcome. There are several techniques to do that when you try to pay attention either on specificity or sensitivity.

But I would suggest to switch the approach and use some Unet-like architecture (in case I understand your question correctly and you are not using it already).

Dear Gunjan,

for deep learning in unbalanced datasets, usually people apply oversampling of the minority class or undersampling of the majority class (both of which are similar concepts). Other people use data augmentation of the minority class. All of these have drawbacks.

However, there is also the possibility to change the loss function such that the majority samples have a smaller learning rate. I believe, such an automatic balancing is implemented by default in the crossentropy/softmax loss function of tensorflow: https://www.tensorflow.org/api_docs/python/tf/nn/weighted_cross_entropy_with_logits

In related research, we have shown that such a balanced training will result in balanced score distributions of your classifier. Usually, the classification of the minority class can be improved, on the cost of decreased performance on the majority class. This might result in a drop in accuracy: Conference Paper MOON: A Mixed Objective Optimization Network for the Recogni...

You can use an weighted cost-function.

Look at this link: https://blog.bigml.com/2017/08/30/using-a-customized-cost-function-to-deal-with-unbalanced-data/

If it's an image data, you could rotate the tumor patches images into say 90 degree or 180 degree to create extra tumor images. Then you need to convert the data into numpy format using Python. Using this method, you're not creating new artificial data but instead, you're just using the same image with a different viewing angle. The last skin lesion image recognition project I did has the same class imbalances problem and I used this method to increases the accuracy of the model.

Class imbalance is a common problem in object detection as well as semantic segmentation. As others have pointed out, one solution is simply to change the sampling rate. Having a 1:5 ratio of positive to negative patches is actually mild in comparison to pedestrian detection. You can fix the ratio of positive to negative patches in each batch at 1:1, reusing your positive patches more often than the negative ones, or you can weight the loss function based on the empirical ratio of samples (but I found this to be inferior).

The method of fixing the ratio of positive to negative examples is actually embedded in the code for state of the art object detectors such as Faster R-CNN to prevent them from predicting every region of interest as background.

We can assign more weight for the minority class and less weight for the majority class to balance the loss.