Williams Index Agreement

Leaving MSRE (Manual Segmentation Guide) is the estimated reference calculated by STAPLE from the three manual segmentations of the sub-regions, with MSREl being the voxeln set titled l in MSRE. First, for each of the twenty themes of each label, the Williams index was calculated using JC as a metric and the eleven Xils input, but only for the sub-regions used for manual tracking. Second, the ASREl was recalculated for each subject, but only on the basis of the sub-regions used for manual follow-up. JC was then measured between the Xil sub-regions and the new ASREl. Third, JC was measured for each advisor between Xil and MSREl (which is defined only for sub-regions used for manual tracing). The average value and standard deviation of the Williams Index and JC score relative to ASRE and MSRE on all subjects are presented in Figure 7. The plots of MDS were also generated on the basis of the average JC table on the twenty themes. Each matrix had 16 × 16 values equal to minus JC in the sub-regions between the 11 segmentations, the three manual segmentations, the MSRE and the new ASRE. The parcels are shown in Figure 8.

The most important measures for MR brain segmentation are the simple match coefficient (SC), the jaccard coefficient (JC) and the dice coefficient (DC). Due to the large number of zeros in our binary cards, a22 is generally much larger than other match numbers and the range of SC values in the actual experience is not wide enough to be properly analyzed. We chose JC as a measure of similarity, since it evaluates only the quantity of horses of the foreground component (Jaccard, 1901): STAPLE was first introduced as a method of assessing the quality of binary segmentation among experts (Warfield et al., 2002a,b), and then was extended to multi-label segmentations (Rohlfing et al., 2003a,b). In this section, we briefly read the multi-label version of STAPLE, a complete description of the method in (Warfield et al., 2004). This algorithm calculates an estimated multi-label reference map from a series of segmentations (rater). Consider a segmented image with n voxels that use one of the possible labels. Let yourself be a × the matrix. Each item describes the probability that I miss a Voxel with t if the real label is t. This matrix resembles the standardized confusion matrix of a Bavarian classificier (Xu et al., 1992) and we will use this terminology for the rest of the paper.

Let`s be the unknown set of all the confusion matrixes that characterize all r-Rater. Let T- (T1, …, Tn)T be a vector representation of unknown actual segmentation and M a matrix n × r whose columns are known segmentations. M are incomplete data and (M, T) complete data. Staple is an estimation process based on the EM algorithm, which can estimate the truth T and the matrix at the same time by maximizing the expectation of the full probability of data log ln-f (M, T|). One of the most interesting features of STAPLE is that it produces performance measures for each segmentation algorithm (confusion matrixes) and a reconciliation of the common agreement, which can be considered an estimated baseline based on the underlying data. For more details on implementation, see Warfield et al. (2004). In terms of the order of priority of the techniques, some observations can be made.

First of all, we have the surprising result that the segmentation of single channels generally fares better than the double channel.

Pin It on Pinterest

Share This