. Author manuscript; available in PMC: 2022 Oct 1.

Published in final edited form as: IEEE Trans Med Imaging. 2021 Sep 30;40(10):2642–2655. doi: 10.1109/TMI.2021.3054817

TABLE IV.

The effect of self-training with different recall values (seen recall, unseen recall and harmonic mean of seen and unseen recall (Harmonic Mean) after the initial multi-view semantic embedding (Init) and after self-training (ST).

	Test Dataset	Seen Recall		Unseen Recall		Harmonic Mean
	Test Dataset	Init	ST	Init	ST	Init	ST
Combination 1	NIH	33.22	30.12	15.3	19.67	20.95	23.8
	NIH-900	46.29	40.72	34.5	45.4	39.54	42.93
	Open-i	52.12	37.74	24.32	41.13	33.16	39.36
	PMC	31.43	25.71	12.12	30.3	17.5	27.82

Combination 2	NIH	7.66	8.68	33.22	31.9	12.45	13.65
	NIH-900	33.17	38.24	29.74	27.96	31.36	32.3
	Open-i	27.94	34.15	31.25	28.75	29.5	31.22
	PMC	20.34	16.95	25	22.73	22.43	19.42

Combination 3	NIH	20	9.28	3.22	8.83	5.54	9.05
	NIH-900	27.33	17.23	6.73	19.81	10.8	18.43
	Open-i	32.74	24.87	6.43	29.08	10.75	26.81
	PMC	15.28	9.72	9.68	22.58	11.85	13.59