Automated Staging Of Mouse Seminiferous Tubules Based On Histomorphometric Analysi

Due to the similarities of spermatogenesis in mammals,earlier studies of male infertility were often modeled in mouse testes.Manually accurate division of the spermatogenesis process from snapshots of the continuous complex dynamic process of spermatogenesis has the potential to more accurately describe the histological and pathological changes of seminiferous duct epithelial cells influenced by various factors.Therefore,the development of a computeraided automatic staging system has the potential to assist pathologists in more accurate staging judgment and lesion diagnoses.The goal of this paper is to construct a computerized spermatogenesis staging(CSS)system for mouse seminiferous tubules based on histopathological image analysis.The main contents and innovations are as follows:(1)Aiming at the problem that the whole slide image(WSI)of mouse testis is large in size and difficult to be processed by conventional methods,this paper proposes a segmentation model based on sliding windows combined with edge overlap prediction.The results show that the sliding window combined with the edge overlap prediction segmentation method proposed in this paper converts the large-size image segmentation problem into a small-size image patch segmentation problem,and effectively solves the difficult problem of segmentation on large-size images.(2)Aiming at the problem that the deep network convolution kernel has a local receptive field,and the local information and global information of the mouse seminiferous tubule are the key features of the staging,this paper proposes an initial staging model of the seminiferous tubule in mice based on multi-scale learning.Compared with the single-scale learning strategy,the multi-scale learning strategy adopted in this paper sends the seminiferous tubule images of different scales into the network for joint training.Furthermore,the multi-scale learning strategy effectively improves the accuracy of the initial staging model by simultaneously extracting the local and global characteristics of the seminiferous tubule.(3)Aiming at the problem of the small number of germ cell labelling samples,this paper proposes a segmentation model of multiple types of germ cells in the seminiferous tubules of mice based on multi-task learning.The results show that outlier aids unrelated to the main task may hinder the segmentation performance of partial tissues when jointly learned.Designing auxiliary tasks that are strongly correlated with the main task is the key to improving the performance of multi-task learning.This requires an in-depth understanding of histopathological images in combination with the experience of pathologists in practice,and the characteristics of pathological images themselves are explored to perform auxiliary tasks.design.The spermatogenesis-related auxiliary task(concentric layer region segmentation)constructed in this paper can improve the performance of germ cell segmentation in the case of limited samples,and has high robustness and generalization.In addition,establishing a multitype cell segmentation model in the mouse seminiferous tubule lays a solid foundation for the subsequent extraction of nuclear features for the more detailed staging of the seminiferous tubules.Based on the image analysis module of the CSS system,the following staging models are respectively developed for the Early Stages,Middle Stages and Late Stages:(4)Aiming at the problem of early(I-V)staging of mouse seminiferous tubules,this paper proposes an automatic staging model of early(I-V)seminiferous ducts in mice based on quantitative image features.In this paper,140-dimensional cell-level morphological features and 64-dimensional cell-level topological features are constructed for Stages I-V.Then,the total 204-dimensional features are filtered by two-sided Mann Whitney U test.Finally,75-dimensional features with significant differences are selected.The results show that the 75-dimensional quantitative image features selected in this paper can effectively differentiate Stages I-III from IV-V,which are also interpretable.It suggests that when the seminiferous tubule transitions from Stages I-III to Stages IV-V,the spermatogonia is in the terminal stage of mitosis,the spermatocyte is undergoing meiosis,the size of various types of germ cells become larger,the color of chromatin in the cell becomes darker,and all the germ cells are in constant motion toward the central region.(5)Aiming at the staging problem of mouse seminiferous tubules(VI-VIII),this paper proposes an automatic staging model of mouse seminiferous tubules(VI-VIII)based on the prior knowledge of pathologists.In this paper,516-dimensional cell-level features are constructed for Stages VI-VIII seminiferous tubules.Another 78-dimensional region-level features are also constructed according to the staging experience of the pathologist,including elongated spermatid orientation and elongated spermatid region texture.In total,594 dimensional features are filtered using the minimum redundancy maximum correlation(m RMR)feature selection algorithm,and 11 dimensional features with the most significant correlation with category labels and the smallest redundancy between them are extracted.The results show that the 11 dimensional quantitative image features can effectively classify Stage VI,Stages VI-m VIII and Stage Late VIII.After feature visualization analysis,it is reflected that the selected features are consistent with the clinical staging experience of pathologists and are interpretable,which verifies the effectiveness of the elongated sperm direction feature designed in this paper.(6)Aiming at the problems of data imbalance and annotation shortage faced by late seminiferous duct(IX-XII)staging in mice,this paper proposes a mouse late seminiferous tubules(IX-XII)staging model based on self-supervised learning.This paper selects a Visual Transformer(Vi T)as the backbone network,adopts a self-supervised learning strategy,and constructs pathology-related self-supervised auxiliary tasks for pre-training.The results show that in the context of a lack of annotations,self-supervised auxiliary tasks designed according to the attributes of histopathological images can effectively improve the staging accuracy of the main task.The application of a self-supervised learning strategy in testicular pathology can potentially relieve the pressure of labelling for pathologists and assist them in more accurate staging.In summary,this paper establishes a computerized spermatogenesis staging system for Stages I-XII of mouse seminiferous tubules based on histomorphometric analysis,which can provide pathologists with quantitative information needed for diagnosis and assist pathologists in staging.The staging models established in this paper all have relatively good interpretability and corroborated with the pathologists’ experience in diagnosis.In addition,the mouse seminiferous tubule staging system was established based on images of H&E staining,which is cheap to use and more generalizable for use in clinical settings.The establishment of a computerized spermatogenesis staging system for mouse seminiferous tubules will not only provide aid to pathologists in staging but can also be integrated with mouse genetic data in future studies to unearth novel comprehensive staging criteria.