Table 1 Comprehensive quantitative comparison of Geo-MIL against state-of-the-art baselines across three public datasets

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

  

WSI Classification (↑)

Segmentation (↑)

Dataset

Method

AUC

Acc

F1

Dice

IoU

TCGA-STAD

AB-MIL4

0.921

0.885

0.881

0.654

0.598

 

TransMIL5

0.953

0.912

0.909

0.702

0.641

 

DSMIL37

0.958

0.915

0.912

0.709

0.650

 

CLAM27

0.949

0.908

0.901

0.695

0.632

 

Patch-WI38

0.905

0.871

0.865

0.631

0.573

 

DTFD-MIL28

0.955

0.913

0.910

0.712

0.655

 

PatchGCN32

0.960

0.921

0.918

0.715

0.658

 

Geo-MIL (Ours)

0.969

0.930

0.927

0.789

0.732

GasHisSDB

AB-MIL4

0.975

0.931

0.929

0.701

0.642

 

TransMIL5

0.988

0.955

0.953

0.758

0.699

 

DSMIL37

0.990

0.960

0.958

0.765

0.708

 

CLAM27

0.985

0.951

0.948

0.749

0.688

 

Patch-WI38

0.969

0.925

0.921

0.685

0.621

 

DTFD-MIL28

0.989

0.958

0.956

0.771

0.715

 

PatchGCN32

0.993

0.968

0.966

0.783

0.728

 

Geo-MIL (Ours)

0.996

0.975

0.973

0.842

0.795

ACDC-GastricDB

AB-MIL4

0.941

0.902

0.899

0.688

0.629

 

TransMIL5

0.965

0.928

0.925

0.731

0.675

 

DSMIL37

0.969

0.933

0.931

0.739

0.684

 

CLAM27

0.962

0.925

0.921

0.725

0.668

 

Patch-WI38

0.933

0.895

0.891

0.669

0.609

 

DTFD-MIL28

0.968

0.931

0.928

0.745

0.691

 

PatchGCN32

0.974

0.941

0.939

0.753

0.699

 

Geo-MIL (Ours)

0.982

0.950

0.948

0.815

0.764

  1. We report performance for WSI Classification (AUC, Acc, F1) and Weakly Supervised Segmentation (Dice, IoU). All values are reported as mean. ↑ indicates that higher is better. The best result per metric within each dataset is shown in bold, and the second-best is underlined.
Back to article page