In practice, this could be really bad for machine learning when applied to cell images (I work for a company that grows modified human cells and collects the images and applies DNNs to them as classifier). DMEM (or FBS, which we use) would effectively cause a systemic bias that would be hard to correct for.