Beyond scalar losses: calibrating segmentation models via gradient vector field surgery

arXiv cs.AI · 6 hours ago

Researchers proposed a method to improve calibration of segmentation models trained with region-based losses like Dice loss by modifying the gradient vector field during training. The technique scales gradient magnitude linearly with prediction error, adding a factor to the loss's partial derivative. Medical imaging segmentation tasks showed improved calibration while maintaining prediction accuracy across 2D and 3D applications.