Table 6 Impact of Macenko stain normalization on cross-dataset generalization performance (Dice Score)

From: Geometric multi-instance learning for weakly supervised gastric cancer segmentation

Method

In-Domain

Cross-Dataset

Cross-Dataset

  

(No Norm)

(+ Stain Norm)

TransMIL

0.758

0.565

0.638 (+7.3 pts)

Geo-MIL (Ours)

0.842

0.741

0.795 (+5.4 pts)

  1. Models were trained on TCGA-STAD and evaluated on the GasHisSDB test set.
  2. The bold values mean the best performanc in the table.
Back to article page