Microscopy is a valuable imaging tool in various biomedical research areas. Recent developments have made high resolution acquisition possible within a relatively short time. State-of-the-art imaging equipment such as serial block-face electron microscopes acquire gigabytes of data in a matter of hours. In order to make these amounts of data manageable, a more data-efficient representation is required. A popular approach for such data efficiency are superpixels which are designed to cluster homogeneous regions without crossing object boundaries. The use of superpixels as a pre-processing step has shown significant improvements in making computationally intensive computer vision analysis algorithms more tractable on large amounts of data. However, microscopy datasets in particular can be degraded by noise and most superpixel algorithms do not take this artifact into account. In this paper, we give a quantitative and qualitative comparison of superpixels generated on original and denoised images. We show that several advanced superpixel techniques are hampered by noise artifacts and require denoising and parameter tuning as a pre-processing step. The evaluation is performed on the Berkeley segmentation dataset as well as on fluorescence and scanning electron microscopy data.