Table 2 Prediction of the performance of different nodes and techniques used for benchmarking breast cancer tumor prediction on the external validation UKA cohort

From: Swarm learning with weak supervision enables automatic breast cancer detection in magnetic resonance imaging

Technique

Local training (AUROC)

Swarm (AUROC)

Centralized Model (AUROC)

Node1 (40%)

Node2 (20%)

Node3 (10%)

3D-ResNet18

0.606 [±0.073]

0.642 [±0.036]

0.534 [±0.038]

0.668 [±0.016]

0.606 [±0.055]

3D-ResNet50

0.610 [±0.089]

0.708 [±0.033]

0.580 [±0.082]

0.750 [±0.019]

0.694 [±0.038]

3D-ResNet101

0.626 [±0.070]

0.664 [±0.171]

0.548 [±0.057]

0.774 [±0.021]

0.742 [±0.026]

3D-DenseNet121

0.622 [±0.028]

0.556 [±0.036]

0.614 [±0.040]

0.656 [±0.038]

0.634 [±0.042]

ViT-MIL

0.604 [±0.038]

0.624 [±0.030]

0.546 [±0.038]

0.660 [±0.041]

0.630 [±0.037]

ViT-LSTM-MIL

0.608 [±0.022]

0.610 [±0.047]

0.532 [±0.036]

0.658 [±0.035]

0.688 [±0.016]

Att-MIL

0.568 [±0.035]

0.540 [±0.020]

0.482 [±0.037]

0.604 [±0.011]

0.558 [±0.039]

2D-ResNet50

0.554 [±0.036]

0.576 [±0.032]

0.538 [±0.035]

0.578 [±0.049]

0.574 [±0.042]

  1. The values of the table represent the mean AUROC, while the errors indicate the standard deviation of AUROC values for each model across five experimental repetitions. The best-performing and lowest performing swarm and centralized models are highlighted in bold.
Back to article page