Table 2 Ablation study investigating the individual contributions of the key components of our Geo-MIL framework

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

 

Model Variant

Dice (↑)

IoU (↑)

AUC (↑)

Full Model

A

Geo-MIL (Full Model)

0.789

0.732

0.969

Ablation of Core Components

B

w/o Graph Representation

0.702

0.641

0.953

C

w/o Topological Gate

0.721

0.665

0.962

D

w/o Dual-Objective (λ = 0)

0.708

0.649

0.961

Analysis of GNN Architecture

E

w/ GNN Layers (L = 1)

0.751

0.695

0.963

F

w/ GNN Layers (L = 5)

0.785

0.728

0.968

G

w/ Neighbors (k = 4)

0.778

0.719

0.966

H

w/ Neighbors (k = 16)

0.783

0.725

0.967

  1. We evaluate the performance on the TCGA-STAD dataset by systematically removing or replacing each component. The removal of any component leads to a notable degradation in performance, particularly in the segmentation metrics, validating our design choices.
Back to article page