Research On Medical Image Recognition Approach

Medical image recognition is a significant part in image-based Computer Aided Diagnosis(CAD)system.In recent years,with the increasingly important role of computers playing in medical diagnosis,the number of digital medical images has increased rapidly and gradually exceeds the head of human abilities.To solve this,CAD system has been widely studied and developed,aiming to solve the drawbacks of manual medical diagnosis,such as time-consuming,subjective,labor-intensive,etc.Then,how to recognize medical images more fully,quickly and accurately,based on limited datasets,becomes a challenging task.Specially speaking,there are two main kinds of medical images recognition methods currently.One is learning effective features by artificial selection at first.Then,doing the recognition using these features.This method has been mature and widely used in CAD system.However,it still need several manual parts and the accuracy also needs to be improved.The other one is using deep neural network directly to extract effective features.In this way,it can effectively reduce the manual parts during the whole process and the accuracy rate is pretty high.However,how to recognize medical images more accurately is still a challenging problem.To address this problem,this paper proposes a series of methods to improve the accuracy of recognition in three medical fields: Human Epithelial-2(HEp-2)cell images(2D)recognition,drug-resistant tuberculosis(3D)recognition and lung cancer(3D)recognition.Specific work can be summarized into the following three aspects:(1)We propose a effective and accurate Human Epithelial-2(HEp-2)cell recognition algorithm.Human Epithelial-2(HEp-2)cell recognition means to classify the images after the segmentation of cell images.The selection of effective features has a great influence on the accuracy of cell recognition,which is particularly significant for those not particularly large datasets.The main method of recognition of HEp-2 cell is using artificial selected feature.A few number of researchers choose deep neural network to extract features directly,such as the use of CNN network to combine feature extraction and classification.However,the overall accuracy of two kinds of methods are not very high.To solve this problem,this paper adopts VGGNET model to extract features automatically and uses SVM to do the classification.The main contents include: 1)proposed an automatic feature extraction and pattern recognition method of HEp-2 cells based on improved VGGNET model;2)improved the VGGNET model in order to be more suitable for datasets which are not particularly large;3)trained SVM to do the classification,which is more effective than using VGGNET model directly;4)evaluated the performance of the proposed method by experiments.(2)We propose an automatic recognition algorithm for tuberculosis drug resistance based on improved depth neural networks.Drug-resistant tuberculosis(TB)has been a persistent death thread of human health for hundreds of years.The increasing emergence of drug resistance and extensively drug-resistant Mycobacterium TB raise researchers' attentions.And how to predict drug-resistant lung TB quickly and effectively has become a big challenge.This paper reviews the major prediction methods of drug-resistant lung tuberculosis appeared in recent years.Specifically,we survey the development of prediction methods of lung TB drug resistance and features selection in different radiological images(CT and X-ray images).Furthermore,we introduce the deep neural network into this field and summarize a framework which is suitable for the prediction process based on previous literatures.In order to improve the accuracy and fully exploiting the information of 3D images,the context information and 3D features are added to the network structure of framework.The test results on the ImageCLEF2017 competition dataset show that the proposed method can accurately predict the resistance to tuberculosis.(3)Application of medical image recognition method based on real CT image.At present,CAD system research is mostly confined to breast and chest lung progressive lesions.CAD research in HEp-2 cell diagnosis,drug-resistant tuberculosis recognition,lung cancer recognition is still very few,and immature.The above two algorithms are both tested on publicly labeled datasets.In order to increase the algorithm's application value,the second part algorithm is applied to the lung cancer CT images provided by the hospital in reality.According to the different final recognition goal,we fine-tuned the algorithm,and the final test results show that the algorithm can get high accuracy in the actual data.