Table 4 Ablation study results on LIDC-IDRI

From: GLANCE: continuous global-local exchange with consensus fusion for robust nodule segmentation

Model variant

Dice (%)

IoU (%)

Full GLANCE (GCT+MRGAM+CSCF)

95.2

90.8

w/o CSCF (no cross-scale fusion)

82.0

69.5

w/o MRGAM (no atrous conv branch)

82.5

70.2

w/o GCT (no transformer branch)

80.0

66.7

w/o GCT, w/ CSCF (two local streams)*

80.4

67.2

w/o MRGAM, w/ CSCF (conv + transformer)*

82.2

69.8

w/o GCT & w/o CSCF (MRGAM only)

80.3

67.1

w/o MRGAM & w/o CSCF (GCT only)

81.0

68.1

w/o GCT & w/o MRGAM (dual plain encoders)

79.5

66.0

w/o GCT & w/o MRGAM & w/o CSCF (baseline U-Net)

79.0

65.3

  1. *These configurations retain two encoder streams. “w/o GCT, w/ CSCF” uses two identical conv-MRGAM streams fused at each scale, while “w/o MRGAM, w/ CSCF” uses one transformer and one plain conv stream.
  2. We report segmentation accuracy as Dice similarity coefficient (DSC) and Intersection-over-Union (IoU). GCT global context transformer, MRGAM multi-receptive grouped atrous mixer, CSCF cross-scale consensus fusion.
Back to article page