On dice loss and non-pixelwise supervision
数学专题报告
报告题目(Title)：On dice loss and non-pixelwise supervision
报告人(Speaker)：Dr. Hoel Kervadec (Erasmus Medisch Centrum)
地点(Place)：后主楼1223
时间(Time)：2023年11月05日（星期日）10:00-11:00
邀请人(Inviter)：段玉萍，王发强
报告摘要
In medical semantic segmentation, dice loss is seen as a standard loss to be combined with standard network architectures. However, despite its widespread use, some of its basic properties have not been well studied, and we show that its inner working mechanism could be simplified. At the same time, the pixel-wise supervision used is akin to "learning by heart" individual pixels, without looking at the "big picture". We investigate promising alternative methods to train segmentation networks not with individual pixels information, but with shape and location information, described in an higher-order fashion.
