Lung vessel segmentation based on random forests

Lung vessel segmentation of computed tomography (CT) images is important in clinical practise and challenging due to difficulties associated with minor size and blurred edges of lung vessels. A vessel segmentation method is proposed for lung images based on a random forest classifier and sparse auto-encoder features. First, the multi-scale representations of lung images are obtained using the Gaussian pyramid. Second, a sparse auto-encoder of three layers is trained using randomly selected patches of these images. Next, the trained weight of the sparse auto-encoder is used as the convolution kernel to extract features of different scale images. Finally, a random forest classifier is exploited to segment the vessels. The proposed method was evaluated on the original and noise-added VESSEL12 dataset that is publicly available. Comparison with some classical methods and existing machine learning methods shows that the proposed method reaches the state-of-the-art accuracy. The results also show that a shallow neural network is a powerful feature extraction tool.