OA142] Validation framework for automated determination of the optimal number of clusters in [F-18]FET-PET brain images

Abstract

Purpose Automated segmentation is used for many applications in medical physics using cluster analysis. However, the definition of the optimal number of clusters is currently not standardized as different cluster validity indices are used for this definition. To meet the general requirements for an adequate cluster validity index, the suggested optimal number of clusters should give a physiologically reasonable segmentation, be independent of the number of investigated voxels and be highly reproducible. To enhance the reproducibility, we investigate the additional requirement ”consistency”, i.e. after removing one cluster from the data the algorithm should suggest the same other clusters. Methods Dynamic [F-18]FET-PET images of eight patients with brain tumors were investigated. A mean activity of (186 ± 7) MBq was injected and a PET measurement in list mode (45 min) was performed. To segment different regions, the k-means clustering algorithm was used. Since k-means needs to specify the number of clusters in advance, WB and I cluster validity indices were used for automated determination of the optimal number of clusters. To check the consistency of the suggested clusters by WB and I indices, one cluster was removed and the remaining voxels were clustered again. The new found clusters were compared with the previous clusters and scored according to 2 factors: (1) production of the expected number of clusters, and (2) the same voxels per cluster before and after removing one cluster. The scores of WB and I indices were compared. Results The optimal numbers of clusters given by WB and I indices were the same for 170/265 slices. McNemar’s test shows that WB and I indices differ significantly (p = 0.025) with respect to the chosen number of optimal clusters. Moreover, the success rates of WB and I indices when removing one cluster were 80% and 73% for factor 1, and 77% and 70% for factor 2, respectively. Conclusions The WB-index outperforms the I-index with regard to consistency of the clustering results. Therefore, the WB-index seems to be the better suited cluster validity index for automated determination of the optimal number of clusters for [F-18]FET-PET brain images.