Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Learning a Loss Function for Segmentation: A Feasibility Study

: Moltz, J.H.; Hänsch, A.; Lassen-Schmidt, B.; Haas, B.; Genghi, A.; Schreier, J.; Morgas, T.; Klein, J.


Institute of Electrical and Electronics Engineers -IEEE-; IEEE Signal Processing Society:
IEEE 17th International Symposium on Biomedical Imaging, ISBI 2020. Symposium Proceedings : 3-7 April 2020, Iowa City, Iowa, USA
Piscataway, NJ: IEEE, 2020
ISBN: 978-1-5386-9330-8
ISBN: 978-1-5386-9331-5
International Symposium on Biomedical Imaging (ISBI) <17, 2020, Iowa City/Iowa>
Fraunhofer MEVIS ()

When training neural networks for segmentation, the Dice loss is typically used. Alternative loss functions could help the networks achieve results with higher user acceptance and lower correction effort, but they cannot be used directly if they are not differentiable. As a solution, we propose to train a regression network to approximate the loss function and combine it with a U-Net to compute the loss during segmentation training. As an example, we introduce the contour Dice coefficient (CDC) that estimates the fraction of contour length that needs correction. Applied to CT bladder segmentation, we show that a weighted combination of Dice and CDC loss improves segmentations compared to using only Dice loss, with regard to both CDC and other metrics.