Research On Segmentation Algorithm Of Human Sperm Head Image Based On Deep Learning

In the direction of artificial intelligence,deep learning is one of the most emerging fields in recent years.In essence,it is a machine learning algorithm that meets a variety of specific needs.Compared with the related technology before it,its effect in language and image processing is far superior to the former.With the continuous updating of highperformance computer chips capable of processing data,deep learning is developing rapidly in computer vision tasks such as image processing,and more and more widely used in the field of medical image processing.And it has vigorously promoted the help of related computer-assisted technologies for the diagnosis of various diseases by professional doctors.It not only improves the work efficiency of doctors,but also makes the diagnosis of various diseases faster and more accurate.With the development of society,due to the increasing pressure of people and the rapid increase in various sources of pollution,male infertility symptoms have become more serious.In clinical practice,semen analysis is necessary to diagnose male infertility and determine the best treatment.Concentration,motility,vitality and the degree of DNA fragmentation of sperm are important indicators analyzed in traditional spermiogram.In addition,another important indicator of the potential fertility of semen samples is the shape of the sperm cells.Therefore,sperm morphology analysis is an important step in the diagnosis of male infertility,and the shape of sperm head is an important indicator in sperm morphology analysis.For this reason,accurate and efficient segmentation of human sperm head is very important for accurate and objective analysis of sperm morphology,as well as for the diagnosis and treatment of male infertility.However,a lot of previous work has shown that the classification and segmentation of sperm head is a challenging task with many aspects.Between the 1970 s and 1990 s,researchers devoted to the analysis of sperm morphology used artificial methods to process sperm morphology.The visual assessment of sperm is also often done manually and is largely dependent on the judgment of a professional physician.Consequently,these traditional approaches has a number of deficiencies,such as time consuming,inaccuracy,subjectivity,non-repeatable,and difficulty in teaching.At present,almost all computer-aided sperm analysis uses different staining procedures to stain sperm,which results in that we can only deal with dead sperm rather than living sperm,so there are serious deficiencies in clinical real-time requirements.In view of the efficient development and huge advantages of deep learning in the medical image processing direction,we apply deep convolutional neural networks(DCNNs)to the segmentation of human sperm heads.The main work of the dissertation is summarized as follows:1.For deep learning,nothing is more important than data.The quantity and quality of data is very important for the improvement of algorithm performance.In practical clinical applications,collecting various types of medical image data is not a simple matter.Although a large amount of data can be collected,not the original data can be used for model training and learning.These data must be identified by relevant professionals before they can be used for deep network learning.For these reasons,the lack of publicly available data sets is one of the greatest difficulties in sperm morphology analysis,as in other medical imaging studies.Therefore,we first collected sperm cell images and build a new unstained dataset that can be used to segment human sperm heads in deep learning algorithms.Our dataset contains 1207 images of sperm cells from more than 20 male infertility patients.2.Due to the shortcomings of traditional methods and full convolutional networks in segmentation,we propose an efficient deep learning algorithm for fully automatic segmentation of human sperm heads based on the U-shaped network structure.We improved the U-shaped network by integrating the stacked residual networks module in its encoding path and decoding path,and use the proposed residual hybrid dilated convolution module to connect the encoding path and decoding path,then we followed the long skip layer in the original network,which finally formed our final deep convolutional neural network.We use our dataset to train our proposed network so that we can segment the sperm head.The final experimental results show that our method is superior to the original U-shaped network in human sperm head segmentation,and it is significantly superior to traditional methods in the clinical application prospect of sperm morphology analysis.3.In view of the fact that the symmetric encoder-deconder segmentation network cannot further fully utilize the semantic information of the image during the image segmentation process.In addition,there are more or less local inconsistencies and semantic inconsistencies in the segmentation results of the existing medical image segmentation models.Therefore,in response to these problems,we added an adversarial learning network on the basis of a dual path segmentation network that can fully extract semantic information,and proposed our dual path segmentation model based on dual discriminator adversarial learning.The proposed network performs sperm head segmentation in an adversarial manner,and we also verified the robustness of our model in the prostate MR dataset.The final results on the two datasets prove that our network has excellent sperm head segmentation performance and good robustness.