Research On Sperm Morphology Analysis Based On Deep Learning

Sperm morphology analysis plays an important role in the assessment of male sperm quality.In the process of traditional sperm morphology analysis,the morphology characteristics of sperm samples need to be observed manually by doctors.This artifact-based analysis of sperm morphology is susceptible to subjective factors such as doctors’ experience.In this paper,deep learning is applied to sperm morphology analysis,and a sperm morphology analysis method is proposed based on segmentation and classification.The main work of this paper is as follows:(1)Most of the public sperm image datasets are small in scale and not fully annotated.In order to perform a complete analysis of sperm morphology,two sets of sperm image datasets,named as Human Sperm Instance Segmentation Dataset(HSIS-Dataset)and Human Sperm Morphology Defect Dataset(HSMD-Dataset),are collected and produced.They are used for sperm instance segmentation and morphology defect classification in our sperm morphology analysis method.The two datasets are completed with the assistance of cooperative laboratory medicine experts and referred to Laboratory Manual for Examination and Handling of Human Semen published by WHO.The two datasets have high scientific research value.(2)In sperm instance segmentation,the convolution enhanced Swin Transformer model combined with Mask R-CNN is used to segment sperm.Based on Swin Transformer,the Patch Merging module with spatial attention and the residual convolution feedforward module are proposed.Spatial attention is used to highlight important feature areas,and convolution is introduced to improve the local modeling ability of the model.In addition,a set of post-processing method is designed to meet the need of sperm morphology defect classification.(3)A classification model of sperm morphology defect based on self-supervised learning and bilinear pooling is proposed.Due to the high difficulty in labeling the classification data of sperm morphology defect,the amount of labeled data is limited.At the same time,a large amount of unlabeled data can be obtained by using the sperm instance segmentation model.Therefore,the self-supervised pre-training method is introduced to use unlabeled data to pre-train the network to improve the effect of downstream supervised learning.In addition,aiming at the problem of small differences in image features of sperm morphology defect,the bilinear pooling is introduced to further improve the effect of sperm morphology defect classification model.(4)Combined with the sperm morphology analysis method described above,a sperm morphology analysis system is designed and implemented.It consists of the client and the server.The client mainly provides functions such as setting relevant information,uploading pictures and generating morphology analysis reports.The server mainly implements the logical processing part of the sperm morphology analysis method,including sperm instance segmentation,post-processing method,morphology defect classification,statistical results generation and other functions.The design of the system takes into account the real needs of the hospital and other use scenarios,which can better assist doctors in sperm morphology analysis.In this paper,deep learning is applied to sperm morphology analysis,and the corresponding deep learning models are proposed according to the characteristics of sperm instance segmentation and sperm morphology defect classification tasks.Based on the proposed models,a sperm morphology analysis system is developed to provide convenience for doctors to conduct sperm morphology analysis.