Table 3 Prediction performance of HRD across three models on test set

From: Predicting homologous recombination deficiency and treatment responses using a computed tomography-based foundation model: a preclinical study

Model

CT Type

AUC

Accuracy

Sensitivity

Specificity

HCR

40 kVp

0.76 [0.69, 0.84]

69% [62%, 77%]

66% [60%, 80%]

71% [57%, 77%]

80 kVp

0.70 [0.61, 0.77]

65% [55%, 68%]

56% [50%, 70%]

75% [54%, 74%]

Combined

0.71 [0.64, 0.79]

65% [57%, 72%]

66% [53%, 74%]

65% [53%, 75%]

sDL

40 kVp

0.85 [0.78, 0.91]

78% [72%, 84%]

73% [62%, 82%]

83% [76%, 91%]

80 kVp

0.77 [0.68, 0.84]

69% [62%, 76%]

73% [68%, 86%]

63% [53%, 72%]

Combined

0.82 [0.73, 0.88]

80% [72%, 85%]

77% [68%, 86%]

83% [73%, 90%]

FM

40 kVp

0.88 [0.82, 0.93]

76% [70%, 86%]

66% [59%, 82%]

87% [76%, 92%]

80 kVp

0.88 [0.83, 0.93]

78% [74%, 86%]

67% [65%, 84%]

89% [78%, 93%]

Combined

0.87 [0.81, 0.92]

77% [70%, 84%]

73% [62%, 81%]

81% [74%, 92%]

Consensus HCR & FM

40 kVp (67%)

0.89 [0.82, 0.95]

84% [75%, 90%]

74% [64%, 90%]

93% [77%, 96%]

80 kVp (62%)

0.90 [0.83, 0.96]

85% [75%, 89%]

70% [63%, 88%]

98% [80%, 96%]

Combined (61%)

0.91 [0.84, 0.96]

85% [75%, 90%]

80% [65%, 89%]

89% [79%, 98%]

Consensus sDL & FM

40 kVp (70%)

0.93 [0.87, 0.98]

89% [81%, 94%]

82% [68%, 90%]

94% [85%, 99%]

80 kVp (65%)

0.91 [0.85, 0.96]

86% [77%, 93%]

81% [68%, 92%]

90% [80%, 97%]

Combined (71%)

0.93 [0.87, 0.98]

90% [84%, 96%]

88% [77%, 96%]

92% [87%, 99%]

  1. HCR handcrafted radiomics, sDL supervised deep learning, FM foundation model, AUC area under the curve.