Materials and Methods
A review was undertaken to identify published studies in the English language that used fMRI as a primary outcome measure of neural responses to visual food cues from 1973 to March 2014. This process is outlined in Figure 1.
Figure 1  Flow diagram of studies included in the review. Initially, electronic databases were searched including: MEDLINE, The Cochrane Library, EMBASE (Excerpta Medica Database), CINAHL (Cumulative Index to Nursing and Allied Health), Informit Health Collection, Proquest, Web of Science, Scopus, and PsycINFO. A pre-determined list of keyword search terms was informed and compiled from a preliminary search of the literature and expanded medical subject headings (MeSH). Keywords were used individually and in combination and included: fMRI, blood oxygen level dependent (BOLD), functional imaging, BOLD signal, BOLD effect, oxyhemoglobin, and deoxyhemoglobin, reward, overeating, addiction, process addiction, food addiction, binge, craving, and dopamine. In addition, electronic searches were supplemented by systematically checking reference lists of relevant publications.
Following the removal of duplicate references, titles and abstracts of identified studies were assessed by two independent reviewers (KP and PS). A predetermined inclusion criterion was applied to determine the study’s eligibility in the review. Studies were included if they investigated an adult population (>18 years of age), used visual food cues, reported weight status, and included fMRI as an outcome measure. Studies reporting a range of BMI categories were included to examine the relationship between neural activation and weight status. Healthy weight individuals were included in the review to act as a comparative group for examining brain activation to visual food cues across all weight status categories. Studies involving participants with a previous or current eating disorder including anorexia nervosa, bulimia nervosa, and binge-eating disorder were excluded, as these participants may have variable responses to food cues that could be attributed to the diagnosis of an eating disorder. Additionally, children or adolescents (<18 years of age); participants with mental and neurological disorders including Prader–Willi Syndrome; pharmacological interventions and fMRI investigations into food intake alone were excluded in keeping with studies included in this review (26–29). If a study included a population meeting the exclusion criteria but reported fMRI outcomes separately for healthy weight and overweight/obese participants, only data on the healthy weight and overweight/obese population was reported in the review. Articles were retrieved for all studies that met the inclusion criteria. If eligibility was unclear, the article was retrieved for further clarification.
Studies were quality checked by two independent reviewers using a standardized 10-question tool (30). The assessed quality criteria included the source of funding, method of sample selection, intervention description, study blinding, and statistical analysis. Four of the quality criteria were designated as “important” and needed to be met to receive a high quality rating. These included: sample selection, comparability of study groups, intervention description, and validity and reliability of outcome measures. An overall classification quality was assigned to each study. Studies were classified as positive quality if >5/10 criteria were satisfied and all important criteria were met. If the majority of criteria were satisfied but at least one of the important criteria was not met, the study was classified as neutral quality. If the majority of the criteria (>5/10) or important questions (≥2/4) were not satisfied, the study was classified as negative quality. Criteria were classified as unclear if the reviewers could not determine whether criteria were met from the detail provided in the published article. Additionally, quality-related fMRI outcomes such as cluster size and volume were extracted and reported in the review. No studies were excluded based on quality ratings. Data were extracted using standardized tables developed for the review. In cases of uncertainty of a study’s inclusion, quality assessment or data extraction were resolved by the consultation of a third independent reviewer until consensus was reached.
Studies were grouped and analyzed by BMI using the World Health Association (WHO) classification, i.e., underweight (<18.49 kg/m2), healthy weight (18.00–24.99 kg/m2), overweight (25.00–29.99 kg/m2), or obese (>30.00 kg/m2) (31). Four groups were created for analysis including: (1) studies that compared healthy weight individuals to overweight/obese individuals; (2) studies investigating individuals pre- and post-weight loss; (3) studies of healthy weight individuals only; (4) studies of overweight/obese individuals only. For the purposes of this study, individuals classified as underweight using the WHO cut points were included in the healthy weight category. Additionally, in studies where BMI spanned a number of categories, the mean BMI was used to classify the study into a specific weight category.

Meta-analysis
To determine the convergence of reported coordinates across studies investigating changes in neural responses pre- and post-weight loss, a meta-analysis was undertaken using the Brainmap GingerALE software1. The inclusion criteria for the meta-analysis were identical to the systematic review criteria. In addition, studies were required to report fMRI outcomes of changes in neural activation to visual food cues pre- to post-weight loss (surgical and behavioral) using either Talairach or Montreal Neurological Institute (MNI) coordinates. Only articles reporting whole brain analysis results were included as region of interest analysis is known to inflate activation findings (32). Papers reporting Talairach coordinates were converted to MNI coordinates prior to analysis using the GingerALE software.
Activation likelihood estimation meta-analysis applies a statistical modeling technique (32) that uses reported brain coordinates and adjusts for between-subject and between-template variance to generate a 3-dimensional Gaussian kernel. Subsequently, a modeled activation (MA) map is created and individual maps are combined to generate an experimental ALE map. The experimental map is tested against an ALE null distribution map, representing the null hypothesis that there is random variation between activation across the meta-analyzed studies, when the within-study variation remains fixed. A random-effects model is applied, which assumes convergence between different studies that is above chance.
A statistical threshold of P < 0.05 False Discovery Rate (FDR), corrected for multiple comparisons and a minimum cluster size of 100 mm3 was set. This is consistent with previous meta-analyses in this area to control for publication bias with respect to the reporting of foci (22–24). Results of meta-analyses are presented using the Mango software package2.