Table 2 Algorithm performance in the Johns Hopkins Comprehensive Stroke Center dataset

From: DeepISLES: a clinically validated ischemic stroke segmentation model from the ISLES'22 challenge

Algorithm

DSC ↑

p-value

F1 ↑

p-value

AVD (ml) ↓

p-value

ALD ↓

p-value

DeepISLES

0.82  ± 0.15

–

0.86  ± 0.33

–

0.84  ± 3.96

–

1.00  ± 2.00

–

[0.45, 0.94]

 

[0.4, 1.00]

 

[0.03, 18.36]

 

[0.00, 9.00]

 

SEALS

0.81  ± 0.16

2.2 × 10−16

0.84  ± 0.33

4.5 × 10−7

0.91  ± 3.95

0.0008

1.00  ± 2.00

0.0026

[0.38, 0.94]

 

[0.4, 1.00]

 

[0.03, 18.62]

 

[0.00, 9.00]

 

NVAUTO

0.82  ± 0.15

1.4 × 10−6

0.80  ± 0.33

2.2 × 10−16

0.84  ± 3.87

0.0072

1.00  ± 3.00

2.2 × 10−16

[0.47, 0.94]

 

[0.4, 1.00]

 

[0.03, 18.50]

 

[0.00, 10.00]

 

SWAN

0.79  ± 0.20

2.2 × 10−16

0.80  ± 0.33

2.2 × 10−16

1.01  ± 4.11

3.0 × 10−9

1.00  ± 3.00

2.5 × 10−7

[0.10, 0.92]

 

[0.29, 1.00]

 

[0.03, 20.65]

 

[0.00, 11.00]

 
  1. DeepISLES significantly outperforms all individual methods and effectively combines their strengths. Values are median  ± interquartile range and [5th, 95th percentile]. Best median values in bold. Wilcoxon signed-rank tests used for comparisons. Source data are provided as a Source Data file.
  2. DSC dice similarity coefficient, F1 lesion-wise F1 score, AVD absolute volume difference, ALD absolute lesion count difference.
Back to article page