Table 1 Unsupervised clustering of cell representations across different methods and datasets

From: VOLTA: an enVironment-aware cOntrastive ceLl represenTation leArning for histopathology

Model

Metric

CoNSeP

NuCLS

PanNuke Breast

PanNuke Colon

Lizard

Oracle

SarcCell

MastCell

MiDOG

Pre-trained ImageNet

AMI

7.3%

9.3%

5.42%

11.21%

6.25%

0.26%

0.8%

0.1 %

13.1%

ARI

7%

7.8%

3.94%

8.21%

4.36%

0.42%

1.8%

0.1 %

5.8 %

Purity

42.7%

56.7%

41.15%

43.93%

50.4%

48.87%

42.0%

58.1%

62.1%

Morphological

AMI

12.7%

21.1%

8.94%

7.88%

13.21%

0.0 %

ARI

1.3%

18.8%

7.28%

6.19%

9.22%

0.0 %

Purity

48.8%

66.1%

47.06%

42.73%

57.5%

58.1%

Manual Features

AMI

9.5%

11.25%

7.86%

10.2%

2.74%

2.9%

2.1 %

6.1%

ARI

6.4%

7.8%

6.53%

3.8%

2.24%

2.1%

4.3 %

7.4%

Purity

45.5%

56.2%

40.37%

52.9%

53.84%

42.7%

62.0%

63.7%

DCAE

AMI

10.1%

8.3%

6.41%

11.43%

4.36%

3.93%

0.0 %

3.5%

ARI

7.3%

7.2%

5.11%

10.01%

2.34%

3.84%

0.0 %

4.3%

Purity

50.5%

56.8%

43.49%

45.18%

49.38%

58.69%

58.1%

60.5%

GAN

AMI

14.8%

14%

6.7%

13.7%

7.5%

4.1%

6.0%

0 %

21.4%

ARI

15.7%

12.6%

4.6%

11.4%

3%

5.8%

5.6%

0 %

27.9%

Purity

58.4%

62%

42.4%

49.6%

48.9%

57.5%

46.0%

58.0%

76.5%

SimCLR

AMI

19.6%

20.1%

10.7%

13.9%

16.5%

12.5%

5.6%

6.2 %

30.2%

ARI

16.7%

22.1%

8.6%

8.9%

11.1%

14.2%

4.5%

8.4 %

30.7%

Purity

57.5%

68.2%

48.3%

40.9%

57.1%

67.5%

45.2%

65.1%

77.7%

DINO

AMI

1.9%

0.6%

0.3%

7.5%

0.4%

0.4%

0.3%

0.0 %

4.9%

ARI

1.7%

0.7%

0.6%

5.5%

0.0%

0.5%

0.8%

0.5 %

6.6%

Purity

1.9%

0.7%

0.5%

7.1%

0.4%

0.6%

41.9%

58.1%

62.9%

VOLTA (w/o env)

AMI

24.2%

22.8%

10.75%

19.5%

10.85%

3.6%

5.3%

0 %

35.5%

ARI

21.7%

24%

7.58%

16.1%

6.2%

2.45%

3.7%

0 %

39.4%

Purity

51.3%

68.3%

46.87%

54.6%

52.66%

54.7%

43.8%

58.1%

81.4%

VOLTA

AMI

25.5%

26.2%

13.8%

22.5%

17.3%

8.05%

4.2%

25.4%

50.4%

ARI

19.3%

27.3%

8.94%

21.8%

11.4%

4.95%

6.7%

33.1%

60.3%

Purity

63.5%

70.3%

47.7%

56.9%

57.9%

59.45%

44.8%

79.0%

88.8%

  1. The best performance is shown in bold.
  2. The baseline models include both morphology-based and state-of-the-art deep learning methods for cell representation. Some of the baseline results are listed as “–" meaning calculation of the feature vectors was not possible due to the limitation of the model on the small-sized cells.