Research And Application Of Mitosis Detection Algorithm Based On Domain Generalization

Mitosis is an indispensable stage in the cell cycle.Accurate detection of mitosis plays a crucial role in diagnosis and treatment in the field of medicine and biology.With the continuous development of deep learning image processing technology,its accuracy in mitosis pathological image detection has surpassed human levels.However,in practical applications,biomedical images come from different hospitals,patients or species,and domain shift problems exist.The feature distributions of images in different domains are different,which can cause the performance of deep learning-based image processing methods to decline or even fail on new domains or datasets,severely limiting the application of deep learning in the field of biomedicine.Additionally,we find that even existing effective target detection algorithms cannot serve as powerful tools for cell biologists to study the process of mitosis.In the field of biomedicine,cell biologists still need to manually measure and record the dynamic changes in spindle length during mitosis under fluorescence microscopy.However,this method is not only time-consuming and labor-intensive,but also has problems of repeatability and accuracy.This article focuses on the domain generalization problem of mitosis detection and proposes a fully automatic fluorescent microscope spindle analysis framework based on this.Specifically,the main research contents are as follows:(1)This article proposes a local feature alignment algorithm and constructs a YOLOv5-LFA target detection domain generalization network.This network adds a local feature alignment branch on the basis of YOLOv5,and maps the local features corresponding to mitosis targets to a representation vector space through a linear mapping function.In this representation vector space,the network uses an improved center loss function to achieve alignment of different domain samples of the same class.The experimental results show that the local representation alignment algorithm helps to improve the generalization performance of the target detection model,and the performance comparison with existing methods also reflects the superiority of the local representation algorithm.(2)This article proposes a fully automatic mitotic spindle analysis framework SpindlesTracker,which mainly consists of three modules:target detection,data association,and skeleton extraction.Firstly,the YOLOX-SP detection network in the target detection module detects spindles in fluorescence microscope images and extracts their boundary boxes and endpoint information.Then,the data association module associates the same spindle instances in the image sequence through the SORT target tracking algorithm.Finally,the skeleton extraction module extracts the skeleton of the spindle through the minimum cost path algorithm,which can then obtain the length of each spindle.Given the current lack of open datasets for spindle analysis,this article contributes a small dataset for budding yeast cells,named S.pombe,and evaluates the performance of SpindlesTracker using this dataset.The results show that the analysis framework performs excellently in the experiment,and the accuracy of each module surpasses existing methods.In summary,this article presents innovative algorithms for research on mitosis in both the medical and biological fields.It provides a new approach to generalize detection algorithms for pathological images and offers an effective method for dynamic detection at the subcellular level.