Пребарување на медицински слики базирано на лонгитудинални податоци за Алцхајмерова болест

Abstract

In this thesis, we performed research towards finding an efficient way to organize and represent the data from the medical cases with the aim to improve their retrieval from the large medical databases. The ultimate goal is to provide semantically and clinically relevant answer from the retrieval system. Particularly, we focused our research on improving the medical case retrieval on the bases of the information extracted from Magnetic Resonance Images (MRI) applied to Alzheimer’s disease. We represented the images with descriptors generated by using the domain knowledge to improve the semantic relevance, precision and efficiency. Considering the nature of the application domain, i.e. the fact that AD is a neurodegenerative disease that progresses over time, we also approached the problem in a longitudinal manner and we represented the medical cases by using longitudinal information. Additionally, we addressed the problem of incomplete data, which is the main challenge regarding the longitudinal data. Considering the limitations of the traditional approach for image representation, we investigated alternative approaches towards getting relevant information and optimal image description by using the domain knowledge. In that direction, we used imaging markers extracted from the regions of interest (ROI) that reflect the brain anatomy, such as volume of the brain structures and cortical thickness to generate the descriptor. Additionally, we explored a representation based on the special pattern of abnormality extracted from MRI. We also included the time component, i.e. we used longitudinal data to represent the patients’ cases. For that purpose, we evaluated feature vectors based on static and dynamic features. Descriptors based on the static measures contain a combination of the volumetric measures of the cortical and sub-cortical regions as well as cortical thickness, extracted from the available images acquired at multiple consecutive time points. On the other hand, dynamic measures, such as rate of change, percent change and symmetrized percent change reflect the severity of the disease and the advance of the degeneration. Additionally, we evaluated a representation comprised of a combination of static and dynamic measures. The experiments were performed with and without quality control (QC) to determine the influence of the errors caused by the automated processing to the results relevance. We also applied feature subset selection to reduce the feature vector dimensionality and to select the most relevant features. Considering the longitudinal data, a key problem that arises is incomplete data, i.e. lack of data for one or more time points. In this PhD thesis, we also proposed a possible solution for this problem. Namely, we suggested to represent the incomplete data with the dynamic features extracted from the available time points because they provide information about the disease progression. Additionally, this strategy provides the same dimension of the descriptors for all patients, regardless the number of the available time points. The proposed techniques were evaluated on a publicly available dataset, provided by the Alzheimer’s Disease Neuroimaging Initiative (ADNI). The evaluation of the proposed strategies showed improvements in comparison to the research already made in the area. Additionally, the obtained results provided answers to some key questions within this area. Thus, the strategies that we used for image representation during the retrieval process in this PhD thesis provided more precise and clinically more relevant, as well as more efficient image retrieval. In fact, the approaches proposed and used in this thesis, directed the retrieval from answering the question “find all cases with similar visual content” (characteristic for the traditional approaches) towards “find all cases with similar structural brain characteristics”, and even more “find all cases with similar brain changes”. This is very important for this research domain. Our future work is directed towards feature learning using deep neural networks. Additionally, we will explore this approach to address the incomplete data as well. Moreover, we will investigate in the direction of dealing with converters, i.e. patients who converted the diagnosis during the period of examination.