Since this question was posted and answered there has been more work in this area. In two papers McGrath and colleagues discuss working directly with the medians and compare this with the transformation methods referenced in other answers. Their papers are "One-sample aggregate data meta-analysis of medians" available here and "Two-sample aggregate data meta-analysis of medians" available here. These are pre-prints and I believe the first at least may be in press.
The two papers are too long to summarise here but in the first paper on summarising in the one-sample case in a nutshell they show that in their simulations working directly with the median is superior to transforming to the mean. They also use an example data-set where the methods give very different results: delay times in care seeking in China. In the second paper things are more complex but it seems that if the majority of the primary studies have presented medians then a median based method is better and if the majority have presented means then the converse applies. They also present methods based on linear quantile mixed models and methods using the quantiles from each primary study to estimate the underlying unknown distribution in each study.
The methods have been provided in an R package metamedian available from CRAN here
Perhaps also worth noting here that they also reference a later paper with a better transformation method by Luo and colleagues entitled "Optimally estimating the sample mean from the sample size, median, mid-range, and/or mid-quartile range" available here