Single photon emission computed tomography (SPECT) is an important diagnostic imaging technique in use for diagnosis and evaluation of cardiac diseases. In gated cardiac SPECT, the data acquisition is synchronized to the electrocardiogram (ECG) signal, and a sequence of 3D images is reconstructed instead of a single static image. Consequently, gated SPECT can provide valuable ventricular function information such as ejection fraction, wall motion, and wall thickening (

Garcia 1996). However, the effectiveness of SPECT often suffers from a number of degrading factors, ranging from decreased image contrast caused by scatter, image artifacts due to attenuation, poor spatial resolution due to distance-dependent blur, to increased noise in gated images.

In recent years there has been considerable interest in development of image processing methods for noise reduction in gated cardiac imaging (

King & Miller 1985,

Narayanan et al. 2000,

Peter et al. 2001,

Brankov et al. 2005,

Klein et al. 1997,

Lalush & Tsui 1998,

Gilland et al. 2002,

Cao et al. 2003,

Gravier et al. 2006). Generally speaking, these methods can be divided into two broad categories: 1) pre- or post-filtering methods, and 2) joint reconstruction methods. Both approaches aim to exploit the statistical correlation (i.e., similarity) among the different gated frames in a cardiac sequence. In the pre- or post-filtering approach, the individual gated frames are reconstructed independently, either preceded by or followed by filtering along the temporal direction (e.g., (

King & Miller 1985,

Narayanan et al. 2000,

Peter et al. 2001,

Brankov et al. 2005)). In the joint reconstruction approach, the gated frames are reconstructed jointly from the entire sequence of data (e.g., (

Klein et al. 1997,

Lalush & Tsui 1998,

Gilland et al. 2002,

Cao et al. 2003,

Gravier et al. 2006)).

In our recent work (

Gravier et al. 2006), we developed a spatio-temporal reconstruction method in which estimated cardiac motion was used to enforce the correlation among the gated frames. We demonstrated the feasibility of this approach based on several quantitative measures using the 4D gated mathematical cardiac-torso (gMCAT) phantom (

Pretorius et al. 1999). Our results showed that incorporation of motion-compensated temporal regularization yields effective noise reduction, resulting in improved perfusion-defect detection in the reconstructed images while avoiding significant cross-frame blurring.

Encouraged by this success, we now further develop and evaluate this approach in a more realistic setting, which is a necessary step toward our ultimate goal of producing a useful clinical tool. In our previous work (

Gravier et al. 2006), we excluded attenuation and scatter so we could isolate and understand the effect of motion compensation by itself. Of course, to be realistic, the algorithm must account for attenuation and scatter. In this paper we now include both, and perform extensive evaluations of the algorithm in this more-realistic situation, including both simulated and patient data.

In this study, we account for the attenuation factor in the reconstruction algorithm by modeling the attenuation map in the system matrix. For scatter compensation, the scatter component is modeled in the likelihood function of the acquired data (as indicated later in

Eq. (1) and thereafter). The resulting 4D reconstruction algorithm will now take into account the following four major degrading factors in gated SPECT: low sensitivity (more severe due to gating), distance-dependent spatial resolution, attenuation, and scatter. This allows us to evaluate the benefit of our 4D approach in a more realistic setting.

Moreover, in our evaluation studies, Monte Carlo simulation (

Ljungberg & Srand 1989) is used to simulate the acquisition data using the more recently developed 4D NURBS-based cardiac-torso (NCAT) phantom (

Segars 2001). Owing to both the complex photon processes inside the phantom and the interactions between the photons and the gamma camera simulated by the Monte Carlo method, this allows for an unbiased evaluation of the reconstruction performance in the presence of a potential mismatch between the actual imaging process and the analytic imaging model used in the reconstruction algorithm.

For the numerical reconstruction we now apply a modified block sequential regularized expectation-maximization (BSREM) algorithm (

Ahn & Fessler 2003), which is globally convergent (

Ahn & Fessler 2003) and much faster than the one-step late (OSL) algorithm used in (

Gravier et al. 2006).

In our evaluation studies, we use the following three quantitative criteria: 1) reconstruction accuracy of the myocardium, quantified by mean squared error (MSE) and bias-variance plot; 2) cardiac perfusion defect detection, measured by a channelized Hotelling observer (CHO) (

Myers & Barett 1987); and 3) time activity curves (TACs) and ejection fraction (EF) of the left ventricle, a clinical measure of left ventricular function. As a preliminary demonstration of the proposed approach, we also provide reconstruction results from a clinical acquisition.

The rest of the paper is organized as follows. The imaging model for gated SPECT and our proposed reconstruction method are described in Sect. II. The evaluation methods are given in Sect. III, and the evaluation results are presented in Sect. IV. Conclusions are given in Sect. V.