Intelligence Recognition And Locolization Of Ovary Common Diseases Based On Ultrasound Images

As a kind of particular apparatus of femininity,ovary is an unignorable part with which a woman can have a healthy body.The variousness,commonly appearance and the potiential ability of deterioration of ovary disease,bring unignorable effect on femininity healthy.Currently,ultrasound examination is a widely used diagnosis method on ovary disease.But using ultrasound imaging equipment to acquire images and make a diagnosis needs doctors having clinical experience for many years,which also have some subjectivity.Our Research is aiming to make use of the state of art of deap learning theory on the field of image recognition and segmentation,so as to apply the approach of artificial intelligence on the clinical diagnosis of ovary diseases.This can realise intelligent recognition of the ovary diseases and the instance segmentation of the infection areas of ovary diseases,without clinical ultrasound doctors’ participating.This also helps the intelligent development on clinical medical science.To make the goals come true,We made two aspect researchs,which are ovary common disease classification based on convolutional neural network and the instance segmentation of the infection areas of ovary common diseases which is based on Mask RCNN.The details are as follows:First,on the aspect of data collecting,collating and pre-processing,We cooperated with doctors from The Fourth Affiliated Hospital of Harbin Medical University.We collected and collated the ovary ultrasound image dataum from sufferers which are totally 2269 images.We created the classification labels of ovary ultrasound image classification dataset and mask labels of infection areas of the ovary common disease.And then,we made two kinds of data pre-processing on ovary ultrasound image dataset which are common methods and fuzzy enhancement and compared the effect on the training of deep learning network without any preprocessing and with the pre-processings.We concluded that common methods and fuzzy enhancement pre-processing can improve the generalization ability of deep learning networks.Second,on the aspect of ovary ultrasound image classification,we studied on the way of feature extraction and the interpretability of the two deep learning network units which are Inception newwork unit and Res Net newwork unit respectively.On the 2269 ovary ultrasound image datum which belong to four kinds of ovary common diseases and normal ovaries,we construted a 23 layers convolutional neural network based on Res Net network unit shortcut connection and compared the realization effect on the mission of ovary ultrasound image dataset classification over the self-defined Res Net unit convolutional neural network,Inception_v3,Inception_Resnet_v2 and the model-agnostic meta-learning model.The self-defined Res Net unit convolutional neural network got a good accurency on ovary ultrasound image dataset classification.Besides,the model size and the forward propagation are small,which can improve the instantaneity on ovary ultrasound image classification mission and save the occupation of video memory.Future more,we reach the highest classification accurency of 81% on ovary ultrasound image dataset with transfer learning method in the end.Third,on the aspect of the insatance segmentation of the infection areas of the ovary common diseases,we used Res Net50 and the Feature Pyrimid Network to extract features on the ultrasound images,used the Region Proposal Network and the shared convolution layers to make objection detection on the infection areas and finally got the semantic segmentation results based on Fully Convolutional Network and the average binary cross-entropy which are all from Mask RCNN.In this way,we completed the mission of instance segmentation of infection areas of ovary common diseases.In the end,we compared the effect on traditional method and the instance segmentation method on the ovary common diseases instance segmentation dataset.