. Author manuscript; available in PMC: 2021 Oct 1.

Published in final edited form as: Med Image Anal. 2020 Jun 20;65:101759. doi: 10.1016/j.media.2020.101759

Table 3.

Results of the experiment on prostate cancer digital pathology classification using different methods. The highest accuracy in each classification task (column) has been highlighted in bold text.

Method	Cancerous vs. benign		High-grade vs. low-grade		Percentage of large classification errors
Method	accuracy	AUC	accuracy	AUC	Percentage of large classification errors
Single pathologist	0.80	0.78	0.65	0.61	0.07
Majority vote	0.86	0.87	0.73	0.74	0.03
STAPLE	0.84	0.86	0.73	0.72	0.03
STAPLE + iMAE loss	0.93	0.91	0.76	0.79	0.03
Minimum-loss label	0.88	0.88	0.80	0.82	0.03
Annotator confusion estimation	0.92	0.93	0.80	0.82	0.01
STAPLE (3–3)	0.86	0.86	0.69	0.70	0.02
STAPLE + iMAE loss (3–3)	0.90	0.88	0.75	0.78	0.02
Annotator confusion estimation (3–3)	0.90	0.88	0.73	0.76	0.03