3Nsemble: Improved Electron Microscopy Image Segmentation Performance with Stacked Generalization

Deep neural networks are widely successful for many tasks of image analysis, including image segmentation. Ensemble models are generally used on deep neural networks not only to enhance the performance but also to improve robustness of predictions. In particular, robustness is currently a limiting factor for image segmentation networks. Here we propose 3Nsemble which uses stacked generalization to improve image segmentation of Electron Microscopy (EM) image data. This research, using neurobiology data, has shown highly accurate automated segmentations of organelles that greatly benefits the study of connectomics and moves us closer to understanding the brain and brain disorders. We compare performance of a trained meta-classifier against simple averaging. The additional costs of training and applying the meta-classifier is outweighed by the benefit of improved performance. The results show improvement in performance metrics with the trained predictions and most notably saw at least a 12% increase in Intersection Over Union (IOU) score