Skip to main content
. Author manuscript; available in PMC: 2010 Aug 5.
Published in final edited form as: Proc SPIE Int Soc Opt Eng. 2010 Mar 12;7623:76231N. doi: 10.1117/12.844182

Figure 2.

Figure 2

Simulations with boundary random raters. Axial sections of the three-dimensional volume show the manually drawn truth model (A) and sample labeling from a single simulated rater (B) alongside STAPLER fused results from 3, 36, and 72 raters producing a total of 3 complete labeled datasets without training data (E-G) and with training data (H-J). Note that boundary errors are generated in three-dimensions, so errors may appear distant from the boundaries in cross-sections. Boundary errors (e.g., arrow in F) increased with decreasing rater overlap. Label inversions (e.g., arrow in G) resulted in very high error with minimal overlap. As with the traditional model (Figure 1), STAPLER enables fusion of label sets when raters provide only partial datasets, but performance suffers with decreasing overlap (C). With the addition of training data (D), STAPLER results in sustained performance improvement even with each rater labeling only a small portion of the dataset.