Maven Loss with AGW-Net for Biomedical Image Segmentation

Traditionally, the dice loss compares the similarity of boundaries between ground truths and predictions. However, the result can be unauthentic when it comes to the situation that both ground truth and predictions are too small. The focal Tversky loss is proposed to address the imbalance between positive and negative samples as well as to contribute a better trade-off between precision and recall. In this paper, we introduce a novel loss function named 'Maven Loss' by considering 'specificity' to handle the issue of data disequilibrium and to help achieve weighing both abilities to correctly segment lesion and non-lesion areas. To evaluate our loss function, we also propose an AGW-Net based on the attention U-Net and W-Net by injecting self-reinforced skip connections. Experiment on ISIC 2018 dataset in which lesions occupy 21.4% on average of the whole images shows that maven loss function and the new network architecture improved IOU and F1-score by 4.9% and 3% compared to the standard attention U-Net, respectively.