Segmenting The Mitochondria From ATUM-SEM Images Using Machine-learning Methods

Mitochondria with double membrane structures are one of the most important organelles in a majority of eukaryotic cells.They carry out all types of important cellular functions including producing the overwhelming majority of cellular ATP needed for endoergonic processes.Meanwhile,they also play an important role in the cellular function,including signaling,controlling the cellular differentiation,growth and death.In addition,they can integrate a variety of specific cell signals to coordinate a variety of complex cell functions,through the dynamic structure changes,such as continuous division and fusion.Recent studies showed that regulation of mitochondrial shape is crucial for cellular physiology,and the localization and morphology of mitochondria have a close connection with neural functionality.Because mitochondria play an important role in physiology,so mitochondrial disorders often lead to a variety of physiological diseases.Therefore,how to automatically detect and segment mitochondria from biological tissue has been found to be very meaningful for the improvement of human health and quality of life.Recently,due to the rapid development of scanning electron microscope(SEM)with its very high resolution,we can look more closely into those fine structures related to cellular function.As a method based on sequence slicing,automated tape-collecting ultramicrotome scanning electron microscopy(ATUM-SEM)adopts the roller designed and belt system method for the ultra thin section,and the backscatter detector of conventional field emission scanning electron for imaging.It is a method which is very suitable for large-scale statistical and analysis of subcellular structure.For the construction of the mitochondria,a complete solution consists of sample preparation,automatic slicing,microscopic imaging,image registration,2D image segmentation and 3D reconstruction.We obtain the first-hand cerebral neuron serial sections by the support of experimental platform of CASIA,and focus attention on the mitochondrial detection and segmentation.From the two aspects,we try to reconstruct the mitochondria from the ATUM-SEM images.Since the mitochondria are widely distributed and have various shapes,and the presence of various subcellular structures produces the sophisticated background.Segmenting the mitochondria from EM images has proven to be a difficult and challenging task.Although the current state-of-the-art algorithms have achieved some promising results,they have demonstrated poor performances on these mitochondria which are in close proximity to vesicles or various membranes.In order to overcome these limitations,this study proposes explicitly modelling the mitochondrial double membrane structures,and acquiring the image edges by way of ridge detection rather than by image gradient.Therefore,we present a coarse-to-fine strategy,which utilizes the Faster R-CNN algorithm for detection,fuses the multi-layer information for the connection across adjacent sections,adopts the total variational model for segmentation.On this base,we utilize the fact that the segmentation results in the continuous layers should maintain consistency,and embed the low-rank property of similar shapes to regularize an active contour model for further optimizing the local misleading segmentation.Finally,we show the 3D visualization in software ImageJ.Experimental results on automated tape-collecting ultramicrotome scanning electron microscopy(ATUM-SEM)images demonstrate the effectiveness of our proposed method.