Biased data, biased AI: deep networks predict the acquisition site of TCGA images

Diagnostic Pathology

Table 1 Comparison of average source site classification performance between KimiaNet’s and DenseNet’s deep features. It suggests that KimiaNet’s deep features contain information pertinent to the tissue source institution

Tissue Source Site Institution	f1-score KimiaNet	f1-score DenseNet	Difference	Number of projects
Indivumed Inc	0.99	0.98	0.01	8
Memorial Sloan Kettering Cancer Center	0.93	0.82	0.11	18
Asterand Bioscience	0.91	0.83	0.08	17
University Health Network	0.91	0.62	0.29	7
Mayo Clinic	0.91	0.70	0.21	9
Barretos Cancer Hospital	0.89	0.61	0.28	7
ILSbio	0.88	0.82	0.06	15
MD Anderson	0.88	0.73	0.15	17
Case Western	0.86	0.72	0.14	2
Johns Hopkins	0.84	0.68	0.16	5
Erasmus MC	0.84	0.67	0.17	4
Roswell Park	0.84	0.68	0.16	18
University of California San Francisco	0.80	0.48	0.32	8
International Genomics Consortium	0.79	0.61	0.18	23
University of Pittsburgh	0.79	0.60	0.19	19
Cureline	0.78	0.56	0.22	10
University of North Carolina	0.75	0.48	0.27	24
Greater Poland Cancer Center	0.70	0.51	0.19	5
Duke University	0.69	0.52	0.17	5
Walter Reed	0.67	0.54	0.13	1
Henry Ford Hospital	0.64	0.32	0.32	2
Christiana Healthcare	0.64	0.50	0.14	17
University of Michigan	0.61	0.29	0.32	6
Emory University	0.41	0.24	0.17	5

ISSN: 1746-1596