关键词: 受限玻尔兹曼机, 肺结节, 良恶性诊断, 计算机辅助诊断

Abstract: For a series of problems in traditional computer-aided diagnosis methods, such as features extraction of lung nodules relying on the manual design, complex operation, low recognition rate, and so on, a diagnosis method of benign and malignant lung nodules based on hybrid restricted Boltzmann machine is proposed. Firstly, multilayer unsupervised convolutional restricted Boltzmann machine is applied in features learning from lung nodules images. Then, these features are used as the input of classification restricted Boltzmann machine to classify benign and malignant lung nodules. In order to avoid the problem of features homogenization during the classification restricted Boltzmann machine training, cross entropy sparse penalty is added to optimize it. Experimental results show that this method can effectively avoid the complexity of manual feature extraction. And it is superior to the traditional diagnostic methods in the accuracy, sensitivity, specificity and area under ROC curve values of classification of benign and malignant lung nodules.

Key words: restricted Boltzmann machine, lung nodules, benign and malignant diagnosis, computer-aided diagnosis