Longitudinal data analysis based on Principal Component Analysis
In clinical studies medical images are now commonly obtained at multiple time points in order to characterize anatomical and functional changes produced by diseases or clinical treatments. However, this set of images, called longitudinal data, leads to a challenge due to “between” and “within” subject variabilities, and the high-dimensional vectors used to stack images. Dimensionality reduction techniques, as principal component analysis (PCA), are thus needed to properly analyze these data. The objective of this presentation is to describe PCA as a tool to reduce dimensionality of high-dimensional data while the main information is preserved. With this, ours was to introduce SVD's definition, properties and geometric interpretation. In addition, we showed the relation between PCA and singular value decomposition (SVD), and the principal component analysis of a synthetic example of a longitudinal data set with 10 subjects. Finally, we want to propose and discuss the extrapolation of the methodology used in the synthetic data as future work in a longitudinal neuroimaging study of Alzheimer's disease.
Expositora: María Clara Salazar Londoño
Fecha y hora: viernes 15 de abril de 2016, 10:00-10:45
Lugar: salón M8-116
Modeling of bladder motion and deformation based on hierarchical and multilevel eigenmodes in prostate cancer radiotherapy
In radiotherapy for prostate cancer the bladder presents the largest inter-fraction shape variations that introduce geometric uncertainties making difficult the proper delivering of the dose. Our interest is then to propose a population model, based on longitudinal data, to estimate bladder motion and deformation between fractions, using only the planning computed tomography (CT) scan as input information. As in previous population models based on principal component analysis (PCA), we use the data to obtain the dominant eigenmodes that describe bladder geometric variations between fractions. However, an important limitation of PCA-based models is that they are not suitable to properly capture the latent structure of multilevel data, such as when subject-level bladders are observed at several time points. With this, ours is hierarchical and multilevel PCA-based models to obtain eigenmodes that separate intra- and inter-patient bladder variability of longitudinal data. In addition, we use mixed-effects models over the data projected in the latent space in order to characterize further intra-patient variability from the total population variance. Based on training data from repeated CT scans, hierarchical and multilevel PCA-based models were thus implemented following dimensionality reduction by means of spherical harmonics (SPHARM). We evaluated the models by means of leave-one-out cross validation on the training data and also using independent data. Probability maps (PMs) were thus generated as predicted regions of probable bladder motion and deformation. These PMs were compared with the observed region using three metrics based on a modified Jaccard index, and mutual information distances. The prediction was compared with two previous population models based on PCA and mixed-effects models.
Expositor: Richard Ríos Patiño
Fecha y hora: viernes 15 de abril de 2016, 11:00-11:45
Lugar: salón M8-116
Ambos seminarios son dirigidos por el profesor Jairo José Espinosa Oviedo
Información adicional: Magda Pinto Vargas, email@example.com , teléfono (+574)4255295
[Boletín UN Investiga 292, 7 de abril de 2016]
[Boletín UN Investiga 293, 14 de abril de 2016]