Tissue-aware dual-attention MIL
The paper introduces CDE-MIL, a weakly supervised multiple instance learning framework for whole slide images. It combines tissue foreground segmentation, feature extraction, and dual-attention aggregation to study colorectal cancer classification without pixel-level annotations.
Three complementary modules
Tissue-Seg
Extract tissue foreground to reduce background interference.
CFE
Combine a variational information bottleneck and top-k selection for feature extraction.
DAM
Aggregate channel-wise and instance-wise attention.
Results reported in the paper
Accuracy (ACC) in the full-data setting is shown below. These values reflect the paper’s dataset splits and experimental conditions, not clinical deployment performance.
| Dataset | Domain | ACC |
|---|---|---|
| TCGA-CRC | Colorectal cancer | 0.9989 |
| CRS | Colorectal cancer | 0.9832 |
| TCGA-STAD | Gastric cancer | 0.9341 |
| CAMELYON16 | Breast cancer lymph node metastasis | 0.9253 |
Author affiliations: Suzhou University of Science and Technology; Shandong University. Summary based on the supplied paper. DOI: 10.1016/j.bspc.2026.110936