Ovarian Follicles Segmentation Based On Deep Learning

Medical image analysis is one of the essential means of clinical auxiliary screening,diagnosis,and classification.In medical images,pathological images can be used to detect diseased cells and tissues for diagnosis or prognosis and to analyze various diseases,such as cancer and multi-organ diseases.The ovary is not only an essential reproductive system of female mammals but also a complex endocrine organ.The ovarian follicle is the cell in a woman’s ovary,suggesting their health.The number of primordial follicles is determined when women are born[21].With the continuous development and maturation of the ovary,follicles are subsequently consumed through recurrent cycles of ovulation.Premature ovarian failure,gene mutation,and drugs will affect the number of follicles.Therefore,by calculating the number of ovarian follicles,we can predict women’s life span and reproductive ability.This paper focuses on the segmentation and counting of ovarian follicles based on deep learning.The main contents are as follows:(1)Optimize the existing deeplabv3+ network and propose the imp-deeplab network,which is more suitable for pathological images.The network adopts a coding and decoding framework and the atrous spatial pyramid pooling(ASPP).We increase its atrous convolution branch and atrous rate.At the same time,low-level information is added.We choose Res Net as the backbone,which has the lower computation.The improved network integrates the low-level and high-level semantic information and can segment the image with high precision.(2)We propose a semantic segmentation system combining attention mechanisms and multi-scale.Firstly,we design a new attention module D-Coordinate,which combines spatial attention and channel attention,and has a good effect.Then,based on HRNet,we design attention HRNet by improving the residual units and adding the attention module D-Coordinate.The network uses multi-scale fusion to ensure that the image maintains high resolution in training and adjusts the weight in the residual module through the attention module to make effective use of high resolution.We tested the segmentation effect of HRNet equipped with different attention modules,and the results show that our attention HRNet is better than the original in pathological images.(3)We apply the improved attention HRNet to ovarian follicles and propose a new automatic segmentation technology system.For the problems of insufficient and unbalanced datasets,we use data augmentation and adjust the weight between different classes.We use the flip,random rotate,and random crop to increase the sets.In order to improve the segmentation accuracy,we adjust the super parameters in the experiment.After segmentation,we add post-processing such as coarse segmentation to ensure the counting.At the same time,we tested the performance of different semantic segmentation networks in ovarian follicle cells.Finally,we compare our method with previous studies.The experimental results show that our method can improve segmentation and counting.