Spatial autocorrelation (Moran’s I) reduces bias of analysis of microscopic images
A general problem of biomedical image analysis is the unbiased selection of the objects of interest. Current methods are mostly based on subjective selection by the observer or on predefined but arbitrary threshold values. These approaches are user dependent, can be difficult to generalize across different samples and have major problems in case of noisy images. Here we provide a user independent spatial correlation analysis (Moran’s I) method that delivers reliable segmentation of objects even with extreme levels of noise and independent of absolute values of image intensity or decisions of the observer. This method does not only consider intensity of pixels but their correlation with their neighbors, too. In model experiments the local Moran’s test selected all pixels of the predefined objects but practically none from the background. Using increasing amount of noise Moran’s test avoided selection of background areas whereas threshold-based segmentations selected more and more of the background pixels. The method was also tested on vGluT2 boutons of thalamic VPM nucleus of rats in the first weeks of postnatal development of sensory ascending pathway. It was capable of delineating boutons on normal and modified images with artificially added noise. Altogether, the Moran’s method did not deliver false positive pixels and it selected true positive pixels similar to either threshold based methods or manual delineation. We propose that this method can also be utilized for image segmentation especially on images with low signal to noise ratio, e.g. intrinsic signal imaging or imaging of calcium sensors.