Institute of Information Theory and Automation

You are here

Bayesian Blind Source Separation in Dynamic Medical Imaging

Ing. Ondřej Tichý
Defense type: 
Date of Event: 
ÚTIA AV ČR, v.v.i., 1. patro, místnost 103
This work is concerned with the blind source separation (BSS) problem in dynamic medical imaging with focus on dynamic planar renal scintigraphy. A common problem of imaging of a three-dimensional object into an image plane is that the signal arises as a superposition of signals from underlaying sources from different depths of a body. The task is to separate individual sources representing functional tissues in medical imaging, i.e. their images and activities over the time. In this work, we study probabilistic models of dynamic image sequences with different assumptions on both, source images and source activities. These assumptions are formalized as a Bayesian model with hierarchical prior and solved by the Variational Bayes (VB) method. The main contribution of this work is introduction of novel models of hierarchical priors for Bayesian BSS, development of BSS algorithms for them, and evaluation of their suitability on clinical data. Common method in this application domain is still manual analysis by an expert. Existing knowledge of the expert in dynamic nuclear medicine was used as inspiration for the proposed hierarchical priors. Two key studied properties of the problem are sparsity of the sources and modeling of each source activity as a convolution of the common input function and a source-specific convolution kernel. As the main contribution, we present several prior models of both, source images and source activities. The key assumption of source images is the sparsity of the signal while it is studied two sparse priors: mixture model and automatic relevance determination (ARD) model. The key assumption of source activities for dynamic medical imaging turned to be assumption that the activity in each source arise as a convolution between common input function and source-specific kernel while various versions of kernels priors are studied. The priors are combined using VB methodology in various manner in order to obtain specific method. The proposed methods are tested together with selected state-of-the-art BSS algorithms on two large datasets from dynamic renal scintigraphy as well as on representative data from other imaging modalities and the significant improvements of separation using proposed methods are demonstrated.
2018-05-03 08:01