Table 7 Preliminary cross-cancer evaluation of Geo-MIL on the TCGA-COAD dataset

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

Dataset

Cancer Type

Dice (↑)

IoU (↑)

TCGA-STAD

Gastric adenocarcinoma

0.789

0.732

TCGA-COAD

Colorectal adenocarcinoma

0.776 ± 0.031

0.718 ± 0.027

  1. The model trained on gastric cancer data (TCGA-STAD) is directly applied to colorectal cancer slides (TCGA-COAD) without fine-tuning. Geo-MIL maintains high Dice and IoU scores, demonstrating its potential generalizability to other epithelial cancers.
Back to article page