Improved pulmonary lung nodules risk stratification in computed tomography images by fusing shape and texture features in a machine‐learning paradigm. Issue 3 (30th December 2020)

• Record Type: Journal Article
• Authors: Sahu, Satya Prakash; Londhe, Narendra D.; Verma, Shrish; Singh, Bikesh K.; Banchhor, Sumit Kumar
• Is Part Of: International journal of imaging systems and technology. Volume 31:Issue 3(2021)
• Page Start: 1503
• Page End: 1518
• Publication Date: 2020-12-30
• Subjects: computed tomography -- computer‐aided diagnosis system -- feature extraction -- feature selection -- nodule segmentation -- pulmonary nodules -- support vector machine; Imaging systems -- Periodicals; Image processing -- Periodicals; 621.367
• DOI: 10.1002/ima.22539
• Languages: English
• ISSNs: 0899-9457

Abstract

Abstract: Lung cancer is one of the most deadly cancer in both men and women. Accurate and early diagnosis of pulmonary lung nodules is critical. This study presents an accurate computer‐aided diagnosis (CADx) system for risk stratification of pulmonary nodules in computed tomography (CT) lung images by fusing shape and texture‐based features in a machine‐learning (ML) based paradigm. A database with 114 (28 high‐risk) patients acquired from Lung Image Database Consortium (LIDC) is used in this study. After nodule segmentation using K‐means clustering, features based on shape and texture attributes are extracted. Seven different filter and wrapper‐based feature selection techniques are used for dominant feature selection. Lastly, the classification of nodules is performed by a support vector machine using six different kernel functions. The classification results are evaluated using 10‐fold cross‐validation and hold‐out data division protocols. The performance of the proposed system is evaluated using accuracy, sensitivity, specificity, and the area under receiver operating characteristics (AUC). Using 30 dominant features from the pool of shape and texture‐based features, the proposed system achieves the highest classification accuracy and AUC of 89% and 0.92, respectively. The proposed ML‐based system showed an improvement in risk stratification accuracy by fusing shape and texture‐based features.