Table 1 Comparison of AUC scores (%) for classification performance on three open-source datasets with varying ratios of annotated samples

From: Enhancing representation in radiography-reports foundation model: a granular alignment algorithm using masked contrastive learning

Method

Zero-shot

CheXpert

NIH

RSNA

  

1%

10%

100%

1%

10%

100%

1%

10%

100%

Ark

× 

–

–

88.7

–

–

82.9

–

–

74.7

M3AE

× 

86.2

87.3

87.9

–

–

–

89.0

90.8

92.3

REFERS

× 

87.2

88.1

88.2

76.7

80.9

84.7

89.4

91.6

92.7

MRM

× 

88.5

88.5

88.7

79.4

84.0

85.9

91.3

92.7

93.3

ConVIRT

✓

85.9

86.8

87.3

66.2

76.6

81.3

77.4

80.1

81.3

GLoRIA

✓

86.6

87.8

88.1

67.1

76.6

81.3

86.1

88.0

88.6

BioViL

✓

–

–

–

69.5

75.3

82.5

88.1

88.4

89.1

MedKLIP

✓

–

–

–

77.2

78.9

83.2

87.3

88.0

89.3

M-FLAG

✓

–

–

–

62.2

71.6

78.7

–

–

–

MaCo (Ours)

✓

88.7

88.7

88.9

79.3

83.8

85.9

91.5

92.7

93.6

  1. Methods with and without zero-shot capabilities have both been included for comprehensive evaluation.