Hybrid-feature-guided lung nodule type classification on CT images

In this paper, we propose a novel classification method for lung nodules from CT images based on hybrid features. Towards nodules of different types, including well-circumscribed, vascularized, juxta-pleural, pleural-tail, as well as ground glass optical (GGO) and non-nodule from CT scans, our method has achieved promising classification results. The proposed method utilizes hybrid descriptors consisting of statistical features from multi-view multi-scale convolutional neural networks (CNNs) and geometrical features from Fisher vector (FV) encodings based on scale-invariant feature transform (SIFT). First, we approximate the nodule radii based on icosahedron sampling and intensity analysis. Then, we apply high frequency content measure analysis to obtain sampling views with more abundant information. After that, based on re-sampled views, we train multi-view multi-scale CNNs to extract statistical features and calculate FV encodings as geometrical features. Finally, we achieve hybrid features by merging statistical and geometrical features based on multiple kernel learning (MKL) and classify nodule types through a multi-class support vector machine. The experiments on LIDC-IDRI and ELCAP have shown that our method has achieved promising results and can be of great assistance for radiologists’ diagnosis of lung cancer in clinical practice.