Table 2 Dice similarity coefficient values in single-modality-missing scenarios for the UCSF-PDGM dataset

From: Unsupervised generative AI for enhancing brain tumor segmentation in multi-center, incomplete real-world data scenarios

Missing sequences

Substitute

DiceWT

DiceTC

DiceET

No missing

/

0.919(0.836–0.952)

0.900(0.733–0.946)

0.797(0.515–0.884)

T1

cpT2

0.901(0.828–0.942)

0.868(0.686–0.938)

0.766(0.485–0.858)

gT1 from T2

0.91(0.832–0.948)*

0.894(0.699–0.946)*

0.774(0.49–0.872)*

T2

cpT1

0.718(0.447–0.809)

0.823(0.427–0.902)

0.69(0.341–0.836)

gT2 from T1

0.818(0.742–0.874)*

0.876(0.607–0.93)*

0.761(0.428–0.855)*

Flair

cpT2

0.707(0.468–0.824)

0.751(0.341–0.865)

0.672(0.382–0.821)

gFlair from T2

0.831(0.659–0.871)*

0.826(0.455–0.912)*

0.757(0.434–0.853)*

T1ce

cpT1

0.853(0.664–0.906)

0.224(0.01–0.492)

0.099(0.008–0.332)

gT1ce from T1

0.898(0.81–0.939)*

0.383(0.016–0.609)*

0.181(0.012–0.372)

DWI

cpT1

0.896(0.795–0.936)

0.886(0.568–0.934)

0.78(0.465–0.87)

gDWI from T1

0.918(0.841–0.951)*

0.9(0.705–0.944)*

0.798(0.478–0.884)*

SWI

cpT2

0.907(0.819–0.944)

0.871(0.609–0.938)

0.782(0.437–0.878)

gSWI from T2

0.911(0.839–0.95)*

0.881(0.67–0.942)*

0.793(0.523–0.877)*

HARDI

cpT1

0.72(0.491–0.858)

0.669(0.292–0.84)

0.474(0.198–0.759)

gHARDI from T1

0.89(0.795–0.932)*

0.86(0.564–0.93)*

0.76(0.462–0.86)*

ASL

cpT1

0.914(0.83–0.949)

0.886(0.715–0.948)

0.784(0.527–0.882)

gASL from T1

0.915(0.835–0.951)*

0.896(0.707–0.947)

0.798(0.541–0.889)

  1. The first column shows missing sequences, and the second column shows the substitute. Values are expressed as medians, with interquartile ranges (25th–75th percentiles) in parentheses. Comparisons were performed using the Wilcoxon signed-rank test when generated images were used to substitute missing sequences versus copied images. *indicates a significant improvement (p < 0.05).
Back to article page