Research On Medical Image Classification Method Based On Restricted Boltzmann Machine

With the development of computer techniques, an increasing number of analytic techniques on medical images by using computers have been developed. Nowadays, applying data-mining methods to the analysis of medical images is becoming popular, and many data-mining methods have been applied on the classification of medical images successfully. These data-mining methods are usually first to extract some descriptions of medical images and then to use statistical learning tools, for example, association rules, decision trees, genetic algorithm, artificial neural network, bayesian network, rough sets and support vector machines(SVMs), to model the dependencies among these features and eventual labels. However, the classification performance of these methods usually have a strong dependence on the statistical features extracted from medical images in advance, and the process of feature extraction is often influenced by many subjective factors such as personal experience.At present, the most popular extractor is called Deep Belief Networks(DBNs) which can learn probabilistic models directly on the raw pixel values. DBNs are multilayer generative models and can also be seen as a directed graphical models consist of a stack of Restricted Boltzmann Machines which are undirected graphical models. For directed graphical models(called belief nets) with dense connectivity between layers, this is usually very difficult because it is generally very hard to infer the posterior distribution over each hidden layer given the input data. So DBN chooses to learn the features one layer at a time using an undirected graphical model called a Restricted Boltzmann Machine(RBM). The RBM has a bipartite connectivity structure that makes it very easy to infer the states of its hidden variables. Once the weights of the RBM have been learned the vectors of hidden feature activations can be used as data for training another RBM that learns a higher layer of features. Hinton has proved that each time an extra layer of learned features is added to a DBN, the new DBN has a variational lower bound on the log probability of the training data that is better than the variational bound for the previous DBN, provided the extra layer is learned in the right way. DBN has been applied a variety of fields because of his capturing the complex higher-order statistical structure that is present in the input data. This paper mainly investigates RBMs which are building blocks of DBNs, and applying their ability of learning features to the analytic of medical images. The main work of this paper is as follows:1. Applying the RBM to the feature learning of medical images.The RBM is building blocks of a DBN, The RBM has a bipartite connectivity structure that makes it very easy to infer the states of its hidden variables and it is an undirected graphical model.RBM is also an effective method to detect features. In this paper, we apply the RBM to the feature learning to medical images to achieve the goal of improving the classification in the process of feature extraction. Then, we use the Bagging-based probabilistic neural network to classify medical images on the features learned by the RBM. The experimental results on the standard data set of studying breast of X-ray images(MIAS)show that Bagging-based PNN performs better in the aspect of classification accuracy rate using the features learned by the RBM than extracted by human beings.2. Applying the DRBM to the classification of medical images.Just using the features learned by the RBM as the input data of other classifier has a certain limitations, for instance, maybe, the format of features learned by the RBM is usuful for the classifier. So, we apply a new method to mammography, which is recently developed in machine learning: Discriminative Restricted Boltzmann Machine which can learn the features automatically from the labeled data and can also perform as a classifier. Discriminative Restricted Boltzmann Machine is a kind of undirected discriminative model. The experimental results on the standard data set of studying breast of X-ray images(MIAS) show that DRBM outperforms Bagging-based PNN based on features learned by the RBM in the aspect of classification accuracy rate.In the end, this thesis lists the current existing problems of the classification of medical images and need to be further research work in the future.