Table 1 Dice similarity coefficient values in all simulated sequence-missing scenarios for the BraTS2019 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

/

0.895 (0.813–0.932)

0.811 (0.578–0.916)

0.790 (0.655–0.883)

T1

empty

0.369 (0.051–0.689)

0.321 (0.025–0.587)

0.204 (0.000–0.641)

cpT2

0.854 (0.704–0.911)

0.712 (0.432–0.877)

0.745 (0.52–0.858)

cpFlair

0.879 (0.757–0.923)

0.798 (0.375–0.907)

0.782 (0.525–0.869)

cpT1ce

0.874 (0.786–0.924)

0.710 (0.528–0.858)

0.763 (0.557–0.871)

gT1

from T2

0.905 (0.774–0.927)*

0.822 (0.545–0.914)*

0.797 (0.617–0.881)*

from Flair

0.902 (0.798–0.926)*

0.810 (0.513–0.918)*

0.800 (0.638–0.878)*

from T1ce

0.906 (0.781–0.927)*

0.800 (0.558–0.913)*

0.803 (0.596–0.877)*

T2

empty

0.520 (0.298–0.745)

0.489 (0.136–0.737)

0.588 (0.122–0.805)

cpT1

0.781 (0.515–0.888)

0.605 (0.225–0.787)

0.719 (0.419–0.835)

cpFlair

0.862 (0.744–0.91)

0.764 (0.505–0.906)

0.773 (0.654–0.879)

cpT1ce

0.851 (0.724–0.902)

0.772 (0.516–0.889)

0.795 (0.65–0.865)

gT2

from T1

0.865 (0.742–0.908)*

0.759 (0.522–0.878)*

0.781 (0.637–0.878)*

from Flair

0.870 (0.775–0.912)

0.803 (0.577–0.902)

0.791 (0.673–0.885)

from T1ce

0.873 (0.763–0.912)

0.788 (0.603–0.903)

0.803 (0.640–0.880)

Flair

empty

0.000 (0.000–0.000)

0.000 (0.000–0.000)

0.000 (0.000–0.000)

cpT1

0.345 (0.147–0.572)

0.542 (0.254–0.824)

0.658 (0.328–0.823)

cpT2

0.554 (0.341–0.689)

0.479 (0.168–0.656)

0.659 (0.371–0.755)

cpT1ce

0.315 (0.153–0.482)

0.661 (0.32–0.846)

0.716 (0.48–0.827)

gFlair

from T1

0.618 (0.417–0.76)*

0.685 (0.439–0.879)*

0.741 (0.625–0.845)*

from T2

0.734 (0.591–0.832)*

0.683 (0.382–0.873)*

0.775 (0.589–0.858)*

from T1ce

0.685 (0.529–0.758)*

0.695 (0.571–0.882)*

0.780 (0.634–0.848)*

T1ce

empty

0.528 (0.284–0.757)

0.016 (0.000–0.182)

0.000 (0.000-0.000)

cpT1

0.849 (0.687–0.914)

0.317 (0.074–0.556)

0.091 (0.0–0.291)

cpT2

0.842 (0.658–0.898)

0.482 (0.273–0.612)

0.248 (0.086–0.378)

cpFlair

0.862 (0.744–0.904)

0.476 (0.335–0.648)

0.296 (0.14–0.397)

gT1ce

from T1

0.894 (0.794–0.917)*

0.295 (0.107–0.488)

0.033 (0.002–0.138)

from T2

0.889 (0.782–0.924) *

0.317 (0.182–0.556)

0.086 (0.011–0.218)

from Flair

0.885 (0.789–0.927)*

0.423 (0.28–0.583)

0.184 (0.038–0.309)

T1 + T1ce

cpT2

0.721 (0.512–0.858)

0.369 (0.181–0.562)

0.182 (0.028–0.324)

cpFlair

0.86 (0.685–0.91)

0.495 (0.315–0.654)

0.313 (0.174–0.458)

gT1_from_T2; gT1ce_from_T2

0.897 (0.756–0.919)*

0.251 (0.129–0.477)

0.075 (0.012–0.136)

gT1_from_Flair; gT1ce_from_Flair

0.884 (0.752–0.922)*

0.404 (0.235–0.555)

0.153 (0.033–0.276)

T2 + Flair

cpT1

0.112 (0.039–0.204)

0.337 (0.035–0.612)

0.44 (0.049–0.715)

cpT1ce

0.229 (0.123–0.351)

0.485 (0.249–0.716)

0.606 (0.407–0.736)

gT2_from_T1; gFlair_from_T1

0.564 (0.37–0.655)*

0.59 (0.365–0.793)*

0.706 (0.555–0.828)*

gT2_from_T1ce; gFlair_from_T1ce

0.571 (0.438–0.674)*

0.659 (0.388–0.803)*

0.718 (0.565–0.832)*

T1 + T2

cpT1ce; cpFlair

0.827 (0.714–0.898)

0.718 (0.442–0.835)

0.751 (0.58–0.85)

gT1_from_T1ce; gT2_from_Flair

0.850 (0.782–0.908)

0.796 (0.577–0.883)*

0.76 (0.624–0.865)

T1 + Flair

cpT1ce; cpT2

0.506 (0.248–0.66)

0.406 (0.167–0.627)

0.6 (0.325–0.745)

gT1_from_T1ce; gFlair_from_T2

0.714 (0.572–0.825)*

0.673 (0.389–0.845)*

0.769 (0.536–0.849)*

T2 + T1ce

cpFlair; cpT1

0.832 (0.66–0.875)

0.357 (0.137–0.564)

0.122 (0.0–0.321)

gT2_from_flair; gT1ce_from_T1

0.854 (0.773–0.892)*

0.33 (0.108–0.482)

0.015 (0.001–0.109)*

T1ce + Flair

cpT1; cpT2

0.436 (0.144–0.615)

0.083 (0.004–0.26)

0.008 (0.0–0.156)

gT1ce_from_T1; gFlair_from_T2

0.71 (0.536–0.81)*

0.113 (0.023–0.377)

0.013 (0.0–0.049)

T1 + T1ce + T2

cpFlair

0.805 (0.639–0.876)

0.481 (0.253–0.678)

0.304 (0.118–0.435)

g_from_Flair

0.844 (0.757–0.89)*

0.4 (0.225–0.535)

0.143 (0.024–0.257)

T2 + Flair + T1ce

cpT1

0.014 (0.001–0.04)

0.0 (0.0–0.001)

0.0 (0.0–0.001)

g_from_T1

0.477 (0.295–0.601)*

0.158 (0.051–0.297)*

0.0 (0.0–0.012)*

T1 + T2 + Flair

cpT1ce

0.156 (0.055–0.263)

0.307 (0.037–0.645)

0.455 (0.158–0.676)

g_from_T1ce

0.569 (0.439–0.671)

0.63 (0.38–0.824)*

0.667 (0.488–0.81)*

T1 + T1ce + Flair

cp T2

0.281 (0.118–0.492)

0.138 (0.037–0.239)

0.043 (0.005–0.123)

g_from_T2

0.691 (0.51–0.778)*

0.128 (0.049–0.35)

0.054 (0.006–0.123)

  1. The first column shows missing sequences, and the second column shows the substitute, e.g. gT1_from_T1ce means generated T1 sequences from T1ce sequences, cp T1ce means directly using T1ce to substitute missing sequences. Values are expressed as medians, with interquartile ranges (25th–75th percentiles) in parentheses. The results for each segmentation type and scenario were compared using the Friedman test or Wilcoxon Signed-Rank Test. Multiple comparisons were adjusted using the Dunn test. *Significant improvement (p < 0.05) was observed when using generated images to substitute missing sequences compared to using copied images.
Back to article page