blob loss: Instance Imbalance Aware Loss Functions for Semantic Segmentation

Deep convolutional neural networks (CNN) have proven to be remarkably effective in semantic segmentation tasks. Most popular loss functions were introduced targeting improved volumetric scores, such as the Dice coefficient (DSC). By design, DSC can tackle class imbalance, however, it does not recognize instance imbalance within a class. As result, large foreground dominate minor instances and still produce satisfactory DSC. Nevertheless, detecting tiny is crucial for many applications, disease monitoring. For example, imperative locate surveil small-scale lesions follow-up of multiple sclerosis patients. We propose novel family functions, blob loss, primarily aimed at maximizing instance-level detection metrics, F1 score sensitivity. Blob designed problems where matters. extensively evaluate DSC-based five complex 3D tasks featuring pronounced heterogeneity terms texture morphology. Compared soft we achieve 5% improvement MS lesions, 3% liver tumor, an average 2% microscopy considering score.