Image quality in MAP SPECT reconstructions

SPECT (Single Photon Emission Tomography) is a noninvasive nuclear medical imaging modality commonly used in the clinic for diagnosis of disease. It is a tomographic modality that displays two-dimensional slices of the three-dimensional spatial distribution of injected radiopharmaceutical within the patient body. The radionuclide in the radiopharmaceutical emits gamma-ray photons that can pass through the body and get detected by a position-sensitive detector. The photons emitted from the radionuclide have a Poisson noise characteristic leading to noisy collected data. Limitations on patient dose mean that few photons are collected and so this photon noise is significant. In addition, the anatomical variability and variations in radiotracer uptake within the bodily anatomical structures is itself a form of noise,"object variability", akin to clutter in radar imaging. A tomographic statistical Bayesian reconstruction algorithm is used to estimate the patient image. The reconstruction algorithm must incorporate an accurate forward model of the imaging system. An inaccurate model results a third type of noise,"model error". All three noise mechanisms propagate into the reconstruction.;In medical imaging "image quality" is assessed by how well information about a task of interest can be obtained from a given image. In my work, the medically relevant task is that of detection of a signal in a noisy background. The signal could be a local increase in radiotracer uptake, for example, indicating the presence of a tumor or other lesion.;My goal in this thesis is to analyze the effects of the three noise sources on detection task performance when the signal is deterministic i.e. its form and location are known. We focus on the use of mathematical model observers that emulate human performance (such as the Channelized Hotelling Observer CHO) in a binary detection task, where the signal is deemed either present or absent. We analyze the CHO and compare it to human performance in an experiment to investigate the optimal (in the sense of maximizing image quality) smoothing in a SPECT MAP (maximum a posteriori) reconstruction. We also observe that in the presence of object variability, CHO and human performance correlate well. We also address another concern regarding the efficacy of a Bayesian reconstruction incorporating an anatomical prior for lesion detection. The anatomical prior uses information regarding the different radiotracer uptake in different organs. Using the CHO we investigate whether this prior information leads to improved detection performance when the only noise source in the data is photon noise. Our results showed no difference in lesion detectability with and without the anatomical information. Finally, we mathematically investigate model error that accounts for noise due to scattered photons in SPECT. We develop theoretical expressions to rapidly compute certain measures of image quality, the reconstructed mean, covariance, and local point-spread function, for SPECT MAP reconstructions. We conclude that this model error adds variance to the reconstructed image above and beyond that due to photon noise.;These investigations are useful in potentially comparing and optimizing both imaging hardware and reconstruction algorithms to achieve better task performance. Even incremental gains in task performance can lead to more favorable diagnostic outcomes for many patients.