Medical Image Analysis And Annotation Based On Deep Learning

With the improvement of living standards,people pay more and more attention to health,but China and even the global medical industry still face many problems.And due to the development of digital technology,huge amounts of medical data are accumulating.So,how to develop these data to alleviate medical-related issues becomes a very meaningful and urgent need.The rise of Deep Learning Networks and their applications in medica image processing point to a feasible path,however,insufficient medical images,missing labels,drifting data distribution,etc,have gradually emerged.Therefore,this thesis studies and proposes a medical image grading model based on deep network transfer and a medical image grading model based on deep anti-transfer network to overcome these obstacles respectively.This thesis first conducts research on the applications of deep learning algorithms in the field of medical imaging,then investigates ensemble learning and transfer learning.It finds that deep learning in the field of medical imaging is facing insufficient data and other issues,but ensemble learning and transfer learning methods show the potential to solve these problems.Therefore,this thesis designs and proposes a medical image classification model based on deep network migration for the problem of ubiquitous medical image shortage.The model is pre-trained on the data unrelated to the target medical images,and then the learned parameters are migrated to the target domain feature extraction network to assist the feature extraction of the target medical images,and then combined with the Random Forest(RF)in Ensemble Learning to obtain the final grading results.For the problem of missing tags,this thesis also proposes a medical image classification model based on deep anti-transfer network.The model combines the grading loss of the auxiliary domain features and the distribution gap loss between the auxiliary domain and the target domain features that are extracted by the same feature extraction network to conduct confrontational guidance on the parameter update of the feature extraction network.This anti-updating operation will enable the extraction network to extract features in both domains that are more favorable to the target domain hierarchy,and achieve the classification of unlabeled images in the target domain finally.The model based on deep network transfer is tested on two types of small-scale medical images of three types of diseases including cataract,diabetic retinopathy and pneumonia.The results show that the model can overcome the impact of target data deficiency,and gain further improvement by combinating with RF,and its application is not limited by the disease type,image type or other similar factors.Experiments of model based on deep anti-transfer network are conducted in different sources of the same diseases and different sources of different diseases scenarios.The results show that the model can classify the unlabeled target data and have some degree of improvement relative to no migration model.