Lung nodule classification using local kernel regression models with out-of-sample extension. (February 2018)

• Record Type: Journal Article
• Authors: Wei, Guohui; Ma, He; Qian, Wei; Han, Fangfang; Jiang, Hongyang; Qi, Shouliang; Qiu, Min
• Is Part Of: Biomedical signal processing and control. Volume 40(2018)
• Page Start: 1
• Page End: 9
• Publication Date: 2018-02
• Subjects: Spectral clustering -- Linear regression -- Kernel trick -- Lung nodules -- Out-of-sample; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2017.08.026
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A novel spectral clustering algorithm LKRME is proposed to classify unlabeled lung nodule. A new Laplacian matrix is constructed to tackle the out-of-sample problem. Performance is evaluated and validated on LIDC-IDRI dataset. Experimental results show the effectiveness of our approach. Abstract: Computer-aided classification is a major research task for computer-aided diagnosis of pulmonary nodules. In radiology domain, labeled data can be expensive to generate. Therefore, in this study, a novel unsupervised spectral clustering algorithm was presented to distinguish benign and malignant nodules. In this algorithm, a new Laplacian matrix was constructed by using local kernel regression models (LKRM) and incorporating a regularization term, the regularization term can tackle the out-of-sample problem. To verify the feasibility of our algorithm, a ground truth dataset was assembled from the LIDC-IDRI database, including 371 benign and 375 malignant lung nodules. All nodules were represented by the texture features, which were computed from the regions of interest (ROIs). Extensive experiments on lung nodules showed that the proposed algorithm not only achieved a higher classification performance than existing popular unsupervised algorithms, but also had superiority comparing to some supervised algorithms (linear discriminant analysis and extreme learning machine).