Calibrated Surrogate Maximization of Dice

Marcus Nordström, Han Bao, Fredrik Löfman, Henrik Hult, Atsuto Maki, Masashi Sugiyama

Abstract: In the medical imaging community, it is increasingly popular to train machine learning models for segmentation problems with objectives based on the soft-Dice surrogate. While experimental studies have showed good performance with respect to Dice, there have also been reports of some issues related to stability. In parallel with these developments, direct optimization of evaluation metrics has also been studied in the context of binary classification. Recently, in this setting, a quasi-concave, lower-bounded and calibrated surrogate for the $$F_1$$ -score has been proposed. In this work, we show how to use this surrogate in the context of segmentation. We then show that it has some better theoretical properties than soft-Dice. Finally, we experimentally compare the new surrogate with soft-Dice on a 3D-segmentation problem and get results indicating that stability is improved. We conclude that the new surrogate, for theoretical and experimental reasons, can be considered a promising alternative to the soft-Dice surrogate.