Objective Assessment Methods for List Mode Imaging System

Abstract

It has been advocated that to properly quantify the imagequality of a medical imaging system, both the task and observer have to be well defined. The objective quality of the image depends on how well the specified observer can extract the desired information to perform the task. Classification or detection tasks are often considered in practice. Two commonly inspected observers are the Hotelling Observer and the ideal observer. The Hotelling observer is optimal among all linear observers in the sense of signal-to-noise ratio, while the performance of the ideal observer sets an upper bound for all other observers by employing the complete statistics of the imaging system and object. Both observers have been extensively studied for traditional medical images that take the form of binned data. Here in this work, we focus on the list mode data, where instead of binning the incident locations, energies, or other formats of attributes, the attributes of each incident are recorded to form a list. This data format has the potential to extract more information from the system and support better performance for clinical tasks. However, the current commonly seen observer constructs can not be applied directly to the list mode data given the unique format. For the Hotelling observer, the conventional implementation operating on binned data normally involves the inversion of covariance matrices and estimation of the difference in means of two vectors. With list mode data, the situation is salvageable by using the attribute list to construct a Poisson point process in attribute space, where instead of the inversion of matrices, the inversion of operators is needed. An example of computing the HOHotelling Observer test statistic on list mode data is provided in Chapter 4. The observer performance is measured on a signal-known-exactly and background-knownstatistically task and compared to the corresponding approximation by use of supervised learning methods on binned data, where a single-layer neural network (SLNN) is used to approximate the HO test statistic. The upper bound set by the ideal observer can be used for the optimization of the imaging system and the performance measure of other observers. However, tracking the performance of the ideal observer has never been easy. For binned data, a Monte Carlo Markov Chain(MCMC) method was proposed. Here in this work, an example of approximating the ideal observer for list mode data given a background known statistically signal known exactly model is provided in Chapter 5. Two methods, the MCMC method and the Taylor series method, are used to perform this task. A comparison is performed with respect to a supervised learning method for binned data, where convolutional neural networks(CNN) are used to approximate the probability of the image data conditioned on the signal present hypothesis, which is a monotone transformation of the IO statistic. The receiver operating characteristic (ROC) curve of the observers is recorded and compared. The results show a superior performance of IO and HO on list mode data compared to its binned form.