Table 2 Results of zero-shot transfer (i.e., performing inference on new data without fine-tuning) on unseen datasets.

From: Universal conditional networks (UniCoN) for multi-age embryonic cartilage segmentation with sparsely annotated data

Model

Trained on all four C57 ages

Mut A E13.5

Mut B E14.5

Mut C E15.5

Mut B E16.5

Avg.

SwinUNETR v2

50.4

71.9

67.1

72.6

65.5

SwinUNETR v2 + ConSA + HDSC

56.9 ↑6.5

78.↑6.2

76.3 ↑9.2

80.9↑8.3

73.1 ↑7.6

Res2UNet*

48.5

69.5

60.6

70.8

62.4

Res2UNet* + ConSA + HDSC

53.7 ↑5.2

76.8 ↑7.3

68.2 ↑7.6

79.9 ↑9.1

69.7 ↑7.3

  1. All the models are trained on the combined training set of the four ages of the C57 dataset (joint training) and applied to the test set volumes belonging to four ages of three different mutations (Mut A, Mut B, and Mut C).
Back to article page