| Mitochondria are one of the most important organelles,which are not only important places for eukaryotes to carry out oxidative metabolism to provide chemical energy to cells,but also important participants in physiological processes such as cell division,differentiation,apoptosis and intracellular information transfer.In clinical practice,a series of diseases are inextricably related to the number,size or structure of mitochondria,and obtaining the number,location,size and shape of mitochondria can be of great help in the diagnosis,etiology tracing and treatment plan determination.Therefore,the design of mitochondrial instance segmentation algorithms with sufficient accuracy is of great importance to both clinical and biological fields.However,in electron microscope images,mitochondria within eukaryotes have characteristics such as variable shapes,different sizes,and complex edges,which pose a considerable challenge to mitochondrial instance segmentation.In order to solve the above problems,based on the single-stage instance segmentation network,this paper deeply studies the effect of the output resolution of the mask on the quality of mitochondria instance segmentation,and on this basis,the attention mechanism is introduced into the feature pyramid structure,which effectively enhances the global awareness of the model and thus obtains a higher quality segmentation mask.The main work is summarized as follows:First,a super-resolution branch-assisted dual-branch mask generation network is designed for the complex features of mitochondrial edges.Rich contextual information is obtained by introducing a densely connected structure with dilation in the super-resolution branch to recover high-frequency features containing rich edge information.And through the feature similarity loss,the mask generation branch is assisted to learn the global relationships in the image reconstruction process to increase the resolution of the network output mask,so as to obtain higher quality segmentation masks.Second,a multi-attention mechanism-guided feature pyramid structure is designed for the characteristics of mitochondria with variable structure and large scale variation.By adding bottom-up information propagation paths to the original feature pyramid structure and using semantic information of different layers in the pyramid structure to guide the bottom-level detail features to propagate upward,it promotes the organic fusion of layer-level features at different scales in the network to enhance the multi-scale representation capability of the network,and then improves the segmentation quality of the network for instances at different scales.Finally,a lightweight non-local attention module is designed to mitigate the influence of the complex environment around the instances in the mitochondria slices on the segmentation results.This module is able to improve the global perception of the network with low time complexity by globally sharing a similarity relationship graph,which in turn makes full use of the surrounding information of the target instance,effectively removes the influence of tiny impurities and similar organelles in the mitochondrial slice on the network segmentation results,and enhances the robustness of the instance segmentation network.In this paper,the effectiveness of the proposed method is verified on a large mitochondrial instance segmentation dataset constructed on the basis of the Mito EM dataset.The experimental results show that,compared with other instance segmentation networks,the method proposed in this paper achieves competitive results on a series of objective evaluation indicators,and can obtain high-quality segmentation masks for instances of different scales in the dataset. |
