OA159] Cluster analysis methods applied to an archive of CT patient dose data

Abstract

Purpose In this study, we investigated the applicability and effectiveness of a data mining technique applied to a CT patient dose archive. In particular, we proposed an automatic extraction of homogeneous groups of CT scans, with isolation of the differences in dose indicators due to the main exposure parameters combinations, compared to those linked to the anatomical variability of the patients. Methods A CT patient dose archive collected with a dose tracking software over the course of a calendar year was considered. It resulted in more than 10,000 examinations and 23,000 scans performed on three equipments, which differ both in terms of the technological generation and the type of diagnostic activity. Hierarchical and not hierarchical clustering methods were applied in the R environment considering different sets of quantitative variables, including only the exposure parameters more influent on image quality and patient dose and not considering the examinations and protocols textual labels. The proper number of cluster for each equipment was determined on the base of the dendrogram and of the within clusters sum of squares. Summary statistics were produced for each cluster and the outliers of the dose indicators distribution were analysed. Results Ward and k-means methods were more effective than other clustering methods. A consistent association between the clusters of homogeneous scanning parameters and the main diagnostic applications was found. Different optimal numbers of clusters, in the range 5–15, were individuated for the three considered equipments. The box and whiskers plots of the dose indicators reflected the homogeneity of the clusters exposure settings and allowed to rapidly identify the protocols with different dose levels. The outliers of the dose indicators distributions reached in some cases the 5% of the scans and many of them were operator’s changes of the standard protocols resulting in a patient dose increase. Conclusions Despite the numerical relevance of the examined datasets, the rapidity of execution and the communicative efficacy of the clustering method outputs are of great interest for the possible implementation of a routine tool for the optimization of the radiation dose in computed tomography.