Multi-view multi-scale CNNs for lung nodule type classification from CT images

• Published in: Pattern recognition 2018-05, Vol.77, p.262-275
• Main Authors: Liu, Xinglong, Hou, Fei, Qin, Hong, Hao, Aimin
• Format: Article
• Language: English
• Subjects: CNNs; Computed tomography; Lung nodule
• ISSN: 0031-3203; 1873-5142
• DOI: 10.1016/j.patcog.2017.12.022

•A comprehensive method for classifying not only solid nodule types such as well-circumscribed and vascularized ones, but also GGO and non-nodule types.•A normalized spherical sampling pattern based on icosahedron and a nodule radius approximation method based on thresholding.•A better view selection method for nodules on CT images based on high frequency content analysis.•A multi-scale multi-view re-sampling and color projection method for n- odules, based on which the CNNs with maximum pooling is trained.•A comprehensive validation on the publicly accessible datasets of LIDC- IDRI and ELCAP. In this paper, we propose a novel convolution neural networks (CNNs) based method for nodule type classification. Compared with classical approaches that are handling four solid nodule types, i.e., well-circumscribed, vascularized, juxta-pleuraland pleural-tail, our method could also achieve competitive classification rates on ground glass optical (GGO) nodules and non-nodules in computed tomography (CT) scans. The proposed method is based on multi-view multi-scale CNNs and comprises four main stages. First, we approximate the spherical surface centered at nodules using icosahedra and capture normalized sampling for CT values on each circular plane at a given maximum radius. Second, intensity analysis is applied based on the sampled values to achieve estimated radius for each nodule. Third, the re-sampling (which is the same as the first step but with estimated radius) is conducted, followed by a high frequency content measure analysis to decide which planes (views) are more abundant in information. Finally, with approximated radius and sorted circular planes, we build nodule captures at sorted scales and views to first pre-train a view independent CNNs model and then train a multi-view CNNs model with maximum pooling. The experimental results on both Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) [1] and Early Lung Cancer Action Program(ELCAP) [2] have shown the promising classification performance even with complex GGO and non-nodule types.