Analyzing the Sensitivity of Affinity Postprocessing Output to Watershed Thresholds in Electron Microscopy Image Segmentation

Abstract

Prior work in the domain of electron microscopy (EM) image segmentation has ad- dressed watershed oversegmentation by using thresholds to preprocess the input to the wa- tershed transform, and hierarchical clustering algorithms to postprocess the output. We initially train convolutional neural networks for affinity detection on these images; then we systematically search for optimal watershed parameters with respect to the quality of initial segmentations and segmentations postprocessed by mean affinity agglomeration, which merges regions in the watershed output based on a local criterion. We illustrate the merits of thresholds that are percentile-based with respect to the model output. We investigate the notion of the existence of an optimal range of percentile parameters across model outputs on a given dataset. We find that mean affinity agglomeration is robust to percentile thresholds within certain ranges, suggesting that in optimizing mean affinity agglomeration output with respect to these parameters, some generality may be afforded with minimal consequence.