Missing Data in Longitudinal Image Retrieval for Alzheimer’s Disease

Abstract

The paper is focused on the missing scans in the context of longitudinal image retrieval for Alzheimer's Disease. Namely, we explore the influence of missing data on the retrieval results when the subjects are represented by the longitudinal changes calculated on the basis of the within-subject template generated using the available time points. To evaluate the effect of the missing scans, we defined two (most characteristic and most common) scenarios, in which missing scans at a specific time point are considered, and one scenario that is based on complete data used as a baseline to compare against. Additionally, we increased the number of patients with missing scans from 10% to 50% and evaluated its impact on the retrieval results. The evaluation showed that from the examined types of feature vectors, concatenated longitudinal changes of the volumes of the cortical and sub-cortical structures are superior and robust. In the case when the dimensionality of the descriptor is an important criterion, we recommend the usage of the percent change or symmetrized percent change of the volumetric measures. Additionally, the influence of the missing scans on the retrieval results is lower when incomplete data occurs in the early time points, rather than in later ones. Moreover, very little or no performance reduction was detected by increasing the number of subjects with missing scans. In general, the evaluation showed very small or no performance degradation in the retrieval process in the scenarios with missing scans, in comparison to the scenario with fully complete data.