An automated method for the extraction of regional data from PET images

Abstract

Manual drawing of regions of interest (ROIs) on brain positron emission tomography (PET) images is labour intensive and subject to intra- and inter-individual variations. To standardize analysis and improve the reproducibility of PET measures, we have developed image analysis software for automated quantification of PET data. The method is based on the individualization of a set of standard ROIs using a magnetic resonance (MR) image co-registered with the PET image. To evaluate the performance of this automated method, the software-based quantification has been compared with conventional manual quantification of PET images obtained using three different PET radiotracers: [¹¹C]-WAY 100635, [¹¹C]-raclopride and [¹¹C]-DASB. Our results show that binding potential estimates obtained using the automated method correlate highly with those obtained by trained raters using manual delineation of ROIs for frontal and temporal cortex, thalamus, and striatum (global intraclass correlation coefficient > 0.8). For the three radioligands, the software yields time–activity data that are similar (within 8%) to those obtained by manual quantification, eliminates investigator-dependent variability, considerably shortens the time required for analysis and thus provides an alternative method for accurate quantification of PET data.

Introduction

Brain images obtained with positron emission tomography (PET) can be analyzed in two different ways: (a) using voxel-based methods or (b) using region-based methods, the latter method being considered superior for data quantification (Hammers et al., 2002). The goal of region-based analysis is the averaging of radioactivity in an anatomic or functional structure, called a region of interest (ROI). Manual techniques for ROI delineation require highly trained personnel and are subject to intra- and inter-operator variations, which can ultimately limit the reproducibility of the results. Additionally, the time and the labour required for manual delineation of ROIs have been increased with the advent of high resolution PET scanners that can produce hundreds of PET slices. To circumvent these limitations, computer-aided methods have been developed to facilitate and improve the reproducibility of the delineation of volumes of interest (VOIs), i.e. set of ROIs describing the single target in a volume space.

Since tracer distribution in PET imaging does not always conform to the simple gray matter/white matter demarcations, or the lobar divisions made on the basis of anatomical divisions (e.g. prefrontal vs. motor cortex) (Evans et al., 1991), direct extraction of ROIs from PET images does not necessarily reflect the ROI's precise anatomical space. While computer vision techniques have been used in some specific situations (Mykkanen et al., 2000, Ohyama et al., 2000, Glatting et al., 2004), indirect determination of ROIs from transformation and registration of atlas-based magnetic resonance (MR) images is the most accepted method to perform region-based analysis of PET images.

Since the earliest work in 1983 (Bajcsy et al., 1983, Bohm et al., 1983), we have seen the development of a number of atlases (Bohm et al., 1991, Greitz et al., 1991, Mazziotta et al., 1995), non-linear image-matching techniques (Collins et al., 1995, Thirion, 1998) of one or more atlases (Hammers et al., 2003) as well as multimodal registration techniques (Woods et al., 1998a, Woods et al., 1998b, Ashburner et al., 1997, Hammers et al., 2002, Studholme et al., 1999). Several automatic methods have been presented for the delineation of ROIs in MR images; however, most of them have not presented an accurate validation to obtain time-activity curves in PET analysis. Two exceptions are the work presented by Yasuno et al. (2002), which we will discuss in detail, and the work of Svarer et al. (2005), which attempts to reduce the individual variability by applying a warping algorithm to several segmented brains to estimate probabilistic ROIs for an individual brain. Yasuno et al. (2002) developed a technique to fit a standard template of ROIs to an individual brain image assisted by a high-resolution reference MR image. This method utilizes computer vision techniques based on the probabilities of gray matter to refine the transformed ROIs. The major limitations of this method, however, are its restricted applicability to sub-cortical regions (particularly the striatum), the template of ROIs expressed in a non-standard brain and its validation using the area under the curve (AUC) of the time–activity data, which may be affected by compensations of excesses and deficiencies of activities.

In the present article, we address these limitations and present the validation of a novel automated method for the extraction of time–activity curves (TAC). First, instead of basing the ROI template on a non-standard space, our approach uses the Montreal Neurological Institute/International Consortium for Brain Mapping (MNI/ICBM) 152 standard brain template, which has the additional benefit of expanding on the number ROIs in the template as well as allowing for a more anatomically valid extension of boundaries pertaining to the respective ROIs. Second, a proper differentiation of gray matter from white matter or cerebrospinal fluid (CSF) is crucial for the accurate delineation of ROIs. This process, also called segmentation, uses a predetermined level of probability of gray matter (threshold). Since the previous method was subject to error, particularly for small ROIs, we present a solution that is based on a fitting function empirically found. Third, one of the key features of an automated ROI program is the establishment of boundaries between adjacent ROIs. In the present approach, we created a natural definition of boundary by using multiple iterations of the morphological dilatation that prevents overlap between neighboring ROIs. Finally, we explored the effect of varying the Full-Width at Half-Maximum (FWHM) of the Gaussian smoothing filter and the use of proton density (PD) weighted MR images to improve the segmentation of the subcortical ROIs.

Our aim is to present a methodology incorporating these corrections that is applicable to cortical as well as subcortical structures such as the caudate and putamen. Our method is validated for its internal consistency and reliability versus trained human raters using PET radioligands with different patterns of brain radioactivity uptake: [¹¹C]-WAY 100635, which is mainly taken up in cortical regions, [¹¹C]-raclopride, which is mainly taken up in the striatum subcortical region, and [¹¹C]-DASB which is taken up in both cortical and subcortical regions.