TABLE III.

Performance of the proposed method in terms of seen precision $\left({\text{PR}}_{\mathcal{S}}\right)$, unseen precision $\left({\text{PR}}_{\mathcal{U}}\right)$, and harmonic mean of seen and unseen precision $\left({\text{PR}}_{\mathcal{H}}\right)$ for different combinations of seen and unseen classes from TABLE I.

	Test Dataset	${\mathbf{\text{Pr}}}_{\mathcal{S}}$	${\mathbf{\text{Pr}}}_{\mathcal{U}}$	${\mathbf{\text{Pr}}}_{\mathcal{H}}$
Combination 1	NIH	37.23	11.27	17.31
	NIH-900	49.64	33.76	40.19
	Open-i	28.17	48.11	35.53
	PMC	34.62	19.61	25.03
Combination 2	NIH	8.52	32.28	13.48
	NIH-900	28.16	36.8	31.9
	Open-i	36.75	20.54	26.35
	PMC	12.99	38.46	19.42
Combination 3	NIH	8.71	9.25	8.97
	NIH-900	22.08	15.16	17.98
	Open-i	40.76	13.99	20.83
	PMC	16.28	11.67	13.59