Research On The Classification System Of Medical Image Based On Semi-supervised Algorithm

Cataract fundus image, a kind of medical image, is often used for the diagnosis of cataract by the doctors clinically. The early identification of cataracts still needs to rely on experienced doctors to diagnose, which will bring great burden to the doctors with the number of cataract patients increasing. So it is very necessary to develop a cataract automatic classification system which will classify the unknown fundus images correctly. The traditional classifier is based on the supervised learning method and the supervised model has a high demand on the number of labeled training samples, but the number of labeled samples in reality is often small. In order to solve this problem, semi-supervised model is used to construct semi-supervised classifier, because semi-supervised can not only use labeled samples, but also can use unlabeled samples. The research work in this paper is mainly reflected in the following aspects:Firstly, the color space transformation is made on the basis of the original image and G-channel image is extracted because of its affluent information. And also the size of the image is normalized.Secondly, the CNN model is used to extract the feature of the image.Based on the characteristics of CNN model, this special extraction method is implicit feature extraction, which is different from the traditional explicit feature extraction method. In this paper, the CNN feature is analyzed and PCA theory is used to reduce the dimensionality of CNN feature. Finally, the feature is used to achieve data visualization.Thirdly, based on the three groups of features, support vector machine and random forest are used to construct the supervised classification models, and the accuracy of the binary-class and four-class are given respectively. It is found that the classification model under CNN feature is better.Fourthly, the label propagation algorithm is combined with the active learning, and label propagation with active learning and Co-forest are used to construct the semi-supervised models. By comparing the effect of semi-supervised classification models based on three different groups of features, we find that the semi-supervised binary classification model under CNN feature reaches 98.3%, meanwhile the number of labeled samples is 150.Fifthly, the PyQt GUI interface is set up under python, and a series of research work in this paper is integrated into the GUI to realize the recognition of cataract fundus images.The purpose of this study is to make use of the semi-supervised model to get a good classification result using fewer labeled samples. The study work can not only solve the problem of shortage of labeled samples in reality, but also play an auxiliary treatment role, facilitating doctors and patients'life.