Table 5.

Performance of homogeneous CXR target tasks using the proposed method and the baseline. Recall that we used three strategies to fine-tune the heads attached to ${\text{UMS-Rep}}_{CXR-Denoising}$. Bold values indicate best performance.

Method	Tuning Strategy	Target Task	Accuracy	F-score	MCC
	Independent	Classification (Bacterial vs Normal)	0.95	0.95	0.90
	Alternating	Classification (Bacterial vs Normal)	0.96	0.96	0.92
	Joint	Classification (Bacterial vs Normal)	0.96	0.95	0.91
	Independent	Classification (Viral vs Normal)	0.95	0.92	0.88
${\text{UMS-Rep}}_{CXR-Denoising}$	Alternating	Classification (Viral vs Normal)	0.96	0.93	0.90
	Joint	Classification (Viral vs Normal)	0.97	0.94	0.92
	Independent	Classification (Bacterial vs Viral)	0.81	0.76	0.56
	Alternating	Classification (Bacterial vs Viral)	0.82	0.77	0.58
	Joint	Classification (Bacterial vs Viral)	0.83	0.80	0.62
Baseline	Separate Model	Classification (Bacterial vs Normal)	0.94	0.94	0.89
	Separate Model	Classification (Viral vs Normal)	0.93	0.90	0.86
	Separate Model	Classification (Bacterial vs Viral)	0.80	0.75	0.54