Poster 286 - ImageSURF: an ImageJ plugin for accurate and unbiased segmentation of fluorescent images

Quantitation of fluorescence images is integral to a wide range of neuroscience research. However, widely used threshold-based methods are sensitive to minor variation in staining and imaging. As an alternative, we have developed ImageSURF (Image Segmentation Using Random Forests), a free open-source ImageJ plugin. ImageSURF uses examples annotated by the user to derive rules to accurately distinguish specific features across large image sets, yielding consistent segmentations regardless of experimental conditions, and unbiased data without requiring experimenter blinding.

ImageSURF has been used for a range of confocal and epifluorescence images, including synapse and microglial markers. For the purposes of evaluation we compared it to optimised thresholding on confocal images of amyloid-beta plaques in a transgenic mouse model of Alzheimer’s disease. Amyloid-beta pathology is difficult to quantify because plaque borders are typically diffuse, and slight variations in thresholds or image brightness can cause large variations in the measured plaque area. We trained ImageSURF using reference segmentations made by human raters, deriving generic rules which could reproduce these reference segmentations when applied to much larger image sets. In thresholding terms, this is equivalent to choosing a global threshold level on the basis of the reference set. In all cases, ImageSURF significantly (p<0.05) outperforms global thresholding, as measured by correlation between the reference segmentations and the outputs of ImageSURF and the best-performing threshold level. On this basis the criteria used to judge a small set of reference images can be applied across the entire image set, yielding an accurate, unbiased quantitation of pathology.