I have been trying to implement this paper and I am very much intrigued. I am working on a medical image problem where I have to segment very small specimens on Whole Slide Images (gigapixel resolution). Therefore my dataset is highly unbalanced and I am having a high false positives rate.
I did my research and found that paper that describes the implementation of Tversky Loss and Focal Tversky Loss. It also describes some modifications to the network architecture which I am postponing for now.
I implemented the loss (Pytorch) and ran some experiments with several alpha/beta combinations. Well, the results are easy to understand: higher alpha results in higher precision and a lower beta increases the recall and pushes the precision down. Basically, what this loss is doing is balancing my recall and precision, only. That is good, I can solve my False Positives issue but since this is a medical problem, a good recall is mandatory. In the paper, the results show that there is an improvement in the Precision/Recall and I cannot understand how is that possible and how I cannot replicate that. I am just weighing false positives and penalizing them, it does not seem enough to improve the model overall.