Table 1 Performance comparison between FLoRIN, NDNT, U-Net models and standard thresholding methods on the SRB and ALRB X-Ray volumes.

From: Flexible Learning-Free Segmentation and Reconstruction of Neural Volumes

Methods

SRB

ALRB

Mean Hausdorff Distance

Runtime (Minutes)

Cell Count

Runtime (Minutes)

FLoRIN

0.9610

14.9111

320

45.17

NDNT*

3.9076

7.8980

253

41.15

NDNT

5.6836

0.2981

220

2.77

Isodata* 45

56.1870

0.9129

9

3.85

Isodata45

20.1399

0.0523

131

3.85

Li* 46

83.5557

6.3752

4

35.84

Li46

18.4453

0.6358

140

1.97

Local50

58.4740

0.4692

7

7

MGAC42

10.5529

237.7531

111

1834.63

Niblack49

43.9941

1.3723

24

6.17

Otsu* 43

100.5323

1.0140

3

4.47

Otsu43

20.5289

0.0576

126

0.48

Otsu 3D43

91.7326

0.0232

0

0.21

Sauvola40

6.2682

1.3383

198

6.22

Triangle* 41

23.8607

0.8932

115

4.44

Triangle41

8.2826

0.0453

203

0.40

Yen* 48

27.7905

0.8809

64

3.77

Yen48

9.6325

0.0460

175

0.46

U-Net32

9.618 ± 1.02 (for 10 slices)

Train: 145.891 Test: 0.049

65 (out of 79)

Not Applicable

3D U-Net33

14.828 ± 2.66 (for 10 slices)

0.0460

53 (out of 79)

Not Applicable

  1. The SRB volume contains 320 total cells. U-Net results are the mean performance of 20 randomly-initialized nets. Asterisks indicate methods applied to small 3D blocks of the volume at a time. U-Net cell counts are reduced due to evaluation on the test set, a subset of 10 images.
Back to article page