Semantic characteristic grading of pulmonary nodules based on deep neural networks

BACKGROUND: Accurate grading of semantic characteristics is helpful for radiologists to determine the probabilities of the likelihood of malignancy of a pulmonary nodule. Nevertheless, because of the complex and varied properties of pulmonary nodules, assessing semantic characteristics (SC) is a difficult task. METHOD: In this paper, we first analyze a set of important semantic characteristics of pulmonary nodules and extract the important SCs relating to pulmonary nodule malignancy by Pearson's correlation approach. Then, we propose three automatic SC grading models based on deep belief network (DBN) and a multi-branch convolutional neural network (CNN) classifier, MBCNN. The first DBN model takes grayscale and binary nodule images as the input, and the second DBN model takes grayscale nodule images and 72 features extracted from pulmonary nodules as the input. RESULTS: Experimental results indicate that our algorithms can achieve satisfying results on semantic characteristic grading. Especially, the MBCNN can obtain higher semantic characteristic grading results with an average accuracy of 89.37%. CONCLUSIONS: Quantitative and automatic grading of semantic characteristics proposed in this paper can assist radiologists effectively assess the likelihood of pulmonary nodules being malignant and further promote the early expectant treatment of malignant nodules.