Table 3 MoE configuration ablation on BraTS21

From: Masked autoencoding, generalizable pretraining, and integrated experts for enhanced glioma segmentation

Configuration

Dice (%)

TP/AP

FLOPs

Latency (ms)

Δ

8 Experts, Top-2

60.87

120M/62M

275G

85

-

4 Experts, Top-2

59.23

91M/62M

275G

84

−1.64

16 Experts, Top-2

61.03

178M/62M

275G

87

+0.16

32 Experts, Top-2

60.94

294M/62M

275G

89

+0.07

8 Experts, Top-1

59.82

120M/47M

261G

81

−1.05

8 Experts, Top-3

60.91

120M/78M

289G

93

+0.04

  1. Metrics measured on H100 GPU. TP/AP=Total/Active Params.
  2. Bold values indicate the best-performing result within each comparison group (i.e., the highest Dice score or the optimal value for the corresponding metric under the same experimental setting).
Back to article page