BioMed Research International has retracted this article. The article was previously submitted to arXiv as: Ondřej Tichý, Václav Šmídl, “Non-parametric Bayesian Models of Response Function in Dynamic Image Sequences,” arxiv, 2015 (https://arxiv.org/abs/1503.05684).

View the full Retraction here.

#### References

- B. Shan, “Estimation of response functions based on variational bayes algorithm in dynamic images sequences,”
*BioMed Research International*, vol. 2016, Article ID 4851401, 9 pages, 2016.

BioMed Research International

Volume 2016, Article ID 4851401, 9 pages

http://dx.doi.org/10.1155/2016/4851401

## Estimation of Response Functions Based on Variational Bayes Algorithm in Dynamic Images Sequences

School of Information Engineering, Chang’an University, Shaanxi 710064, China

Received 16 April 2016; Revised 8 June 2016; Accepted 17 July 2016

Academic Editor: Zexuan Ji

Copyright © 2016 Bowei Shan. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

#### Abstract

We proposed a nonparametric Bayesian model based on variational Bayes algorithm to estimate the response functions in dynamic medical imaging. In dynamic renal scintigraphy, the impulse response or retention functions are rather complicated and finding a suitable parametric form is problematic. In this paper, we estimated the response functions using nonparametric Bayesian priors. These priors were designed to favor desirable properties of the functions, such as sparsity or smoothness. These assumptions were used within hierarchical priors of the variational Bayes algorithm. We performed our algorithm on the real online dataset of dynamic renal scintigraphy. The results demonstrated that this algorithm improved the estimation of response functions with nonparametric priors.

#### 1. Introduction

Highly rapid development of machine learning technique offers an opportunity to obtain information about organ function from dynamic medical images, instead of invasive intervention. The unknown input function can be obtained by deconvolution of the organ time-activity curve and organ response function. Typically, both the input function and the response functions are unknown. Moreover, the time-activity curves are also not directly observed since the recorded images are observed as superposition of multiple signals. Analysis of the dynamic image sequences thus require to separate the original sources images and their weights over the time forming the time-activity curves (TACs). The TACs are then decomposed into input function and response functions. Success of the procedure is dependent on the model of the image sequence.

The common model for dynamic image sequences is the factor analysis model [1], which assumes linear combination of the source images and TACs. Another common model is that TAC arises as a convolution of common input function and source-specific kernel [2, 3]. The common input function is typically the original signal from the blood and the role of convolution kernels varies from application area: impulse response or retention function in dynamic renal scintigraphy [4]. In this paper, we will refer to the source kernels as the response functions; however other interpretations are also possible.

Analysis of the dynamic image sequences can be done with supervision of experienced physician or technician, who follows recommended guidelines and uses medical knowledge. However, we aim at fully automated approach where the analysis fully depends on the used model. The most sensitive parameter of the analysis is the model of the response functions (i.e., the convolution kernels). Many parametric models of response functions have been proposed, including the exponential model [5] or piecewise linear model [6, 7]. An obvious disadvantage of the approach is that the real response function may differ from the assumed parametric models. Therefore, more flexible classes of models based on nonparametric ideas were proposed such as averaging over region [8], temporal regularization using finite impulse response filters [9], or free-form response functions using automatic relevance determination principle in [10].

In this paper, we will study the probabilistic models of response functions using Bayesian methodology within the general blind source separation model [11]. The Bayesian approach was chosen for its inference flexibility and for its ability to incorporate prior information of models [12, 13]. We will formulate the prior model for general blind source separation problem with deconvolution [10] where the hierarchical structure of the model allows us to study various versions of prior models of response functions. Specifically, we design different prior models of the response functions with more parameters than the number of points in the unknown response function. The challenge is to regularize the estimation procedure such that all parameters are estimated from the observed data. We will use the approximate Bayesian approach known as the variational Bayes method [14]. The resulting algorithms are tested on synthetic as well as on real datasets.

#### 2. Probabilistic Model of Image Sequences

A probabilistic model of image sequences is introduced in this section. Estimation of the model parameters yields an algorithm for Blind Source Separation and Deconvolution. Prior models of all parameters except for the response functions are described here while the priors for the response functions will be studied in detail in the next section.

##### 2.1. Model of Observation

Each recorded image is stored as a column vector , , where is the total number of recorded images. Each vector is supposed to be an observation of a superposition of source images , , stored again columnwise. The source images are weighed by their specific activities in time denoted as . Formally,where is the noise of the observation, is the matrix composed of source images as its columns , and symbol denotes transposition of a vector or a matrix. Equation (1) can be rewritten in the matrix form. Suppose that the observation matrix and the matrix with TACs in its columns . Note that we will use the bar symbol, , to distinguish the th row of matrix , while will be used to denote the th column. Then, (1) can be rewritten into the matrix form as

The tracer dynamics in each compartment is commonly described as convolution of common input function, vector , and source-specific response function (convolution kernel, mathematically), vector , [5, 6, 15]. Using convolution assumption, each TAC can be rewritten aswhere the matrix is composed of elements of input function asSuppose that the aggregation of response function . Then, and model (2) can be rewritten as

The task of subsequent analysis is to estimate the matrices and and the vector from the data matrix .

##### 2.2. Noise Model

We assume that the noise has homogeneous Gaussian distribution with zero mean and unknown precision parameter , . Then, the data model (2) can be rewritten aswhere symbol denotes Gaussian distribution and is identity matrix of the size given in its subscript. Since all unknown parameters must have their prior distribution in the variational Bayes methodology, the precision parameter (inverse variance) has a conjugate prior in the form of the Gamma distributionwith chosen constants shape parameter and scale parameter , due to the homogeneous noise model.

##### 2.3. Probabilistic Model of Source Images

The only assumption on source images is that they are sparse; that is, only some pixels of source images are nonzeros. The sparsity is achieved using prior model that favors sparse solution depending on data [16]. We will employ the automatic relevance determination (ARD) principle [17] based on joint estimation of the parameter of interest together with its unknown precision. Specifically, each pixel of each source image has Gaussian prior truncated to positive values (see Appendix A.1) with unknown precision parameter which is supposed to have conjugate Gamma prior asfor , , and , are chosen constants. The precisions form the matrix of the same size as .

##### 2.4. Probabilistic Model of Input Function

The input function is assumed to be a positive vector; hence, it will be modeled as truncated Gaussian distribution to positive values with scaling parameter as where denotes zeros matrix of the given size and , are chosen constants.

##### 2.5. Models of Response Functions

So far, we have formulated the prior models for source images and input function from decomposition of the matrix . The task of this paper is to propose and study prior models for response functions as illustrated in Figure 1. Different choices of the priors on the response functions have strong influence on the results of the analysis which will be studied in the next section.