Premium
Pulmonary lesion classification from endobronchial ultrasonography images using adaptive weighted‐sum of the upper and lower triangular gray‐level co‐occurrence matrix
Author(s) -
Khomkham Banphatree,
Lipikorn Rajalida
Publication year - 2021
Publication title -
international journal of imaging systems and technology
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.359
H-Index - 47
eISSN - 1098-1098
pISSN - 0899-9457
DOI - 10.1002/ima.22517
Subject(s) - percentile , pattern recognition (psychology) , support vector machine , artificial intelligence , computer science , gray level , mathematics , pixel , statistics
Abstract Visual classification of pulmonary lesions from endobronchial ultrasonography (EBUS) images is performed by radiologists; therefore, results can be subjective. Here, two robust features, called the adaptive weighted‐sum of the upper triangular gray‐level co‐occurrence matrix (GLCM) and the adaptive weighted‐sum of the lower triangular GLCM (AWSL), were combined with 22 other standard features and used as initial input data to assist radiologists. The proposed method integrated the k th percentile of the sum of intensities, a genetic algorithm (GA), and support vector machine (SVM) to classify a lesion, and then applied the k th percentile of the sum of intensities to select the optimal window of interest (WOI) where all the features are extracted. After feature extraction, a GA was used to select only relevant features that were then forwarded to SVM to classify the lesion. Efficiency of the proposed features and the proposed method was evaluated using a dataset of 89 EBUS images with 10‐fold cross‐validation. Optimal classification results were obtained using 16 selected features from the WOI at the fifth percentile with accuracy, sensitivity, specificity, and precision at 86.52%, 87.27%, 85.29%, and 90.57%, respectively. Among the 16 selected features, six were from the proposed features. The proposed method was compared with other existing methods. Results revealed that the proposed features together with the proposed method significantly improved the classification performance of pulmonary lessons, especially for small datasets.