Segmentation and classification of ground glass nodule on CT images

The symptoms of lung cancer mainly manifest as lung nodules. The recognition and diagnosis of a ground glass nodule (GGN) are relatively difficult, and its image features are not easy to extract. To improve the accuracy of segmentation and invasive classification of GGN, a region adaptive Markov random field (MRF) model and a two-channel integrated network based on densely connected convolutional neural networks (DenseNet) are developed in this paper. First, the lung parenchyma is segmented coarsely, and the corresponding contour is repaired. Afterwards, the fuzzy C-means (FCM) clustering algorithm and a simple MRF model for the coarse segmentation of GGN are applied. Then, the region adaptive MRF model for the fine segmentation of GGN is used. Moreover, the two-channel integrated network based on DenseNet is used for GGN invasive classification. Results show that the average value of the overlapping area ratio of the segmentation result to that obtained by one physician was 0.9144, and the accuracy of the proposed GGN invasive classification model, the specificity, sensitivity, and the area under curve were 92.553%, 87.500%, 97.826%, and 0.9715, respectively. The proposed segmentation method can segment GGN more accurately, and the corresponding GGN invasive classification model achieved satisfying classification performance.