The primary set of meta-analyses in Neurosynth are based on the occurrence of terms in the abstracts of published neuroimaging articles. At present, the database contains 1335 term-based meta-analyses ranging from action observation to working memory.
The topic-based meta-analyses on Neurosynth represent an effort to move beyond individual term occurrences by using a standard topic modeling approach (Latent Dirichlet allocation) to identify latent topics in the abstracts of articles in the database. For a more detailed explanation, see Poldrack et al (2012). Here we provide several sets of topics estimated at different resolutions, ranging from relatively coarse (50 topics) to more fine-grained (200 topics).