Adaptive weighted aggregation in Group Improvised Harmony Search for lung nodule classification. Issue 2 (3rd March 2020)

• Record Type: Journal Article
• Authors: Kar, Subhajit; Das Sharma, Kaushik; Maitra, Madhubanti
• Is Part Of: Journal of experimental & theoretical artificial intelligence. Volume 32:Issue 2(2020)
• Page Start: 219
• Page End: 242
• Publication Date: 2020-03-03
• Subjects: Evolutionary optimisation -- adaptive weight selection -- Group Improvised Harmony Search (GrIHS) -- adaptive K-nearest neighborhood (AKNN) classifier -- feature subset selection -- lung nodule classification; Artificial intelligence -- Periodicals; 006.3
• DOI: 10.1080/0952813X.2019.1647561
• Languages: English
• ISSNs: 0952-813X

Abstract

ABSTRACT: This work, primarily, addresses the problem of automated diagnosis of lung cancer by classifying malignant nodules present in the lung, if any. To achieve the goal, we have posed a weighted dual objective optimisation problem so as to reduce the feature subset required for automated classification of malignant lung nodules and at the same time, we endeavour to increase the classification accuracy to minimise the false impression. The values of the respective weights associated with the discriminating feature subset and the classification accuracy have been evaluated in such a way that the accuracy of nodule classification increases with the help of a small discriminating feature subset. The strategic weight selection methodology, presented in this work, helps attain an optimal combination of these two objectives. As a solution methodology, subsequently, we propose a new adaptive weighted aggregation strategy based on an evolutionary optimisation technique, popularly known as Group Improvised Harmony Search (GrIHS). An adaptive K NN classifier has also been embedded with the GrIHS method to determine the optimal number of neighbourhoods, in the search space, that would further aid in increasing the nodule classification accuracy. The proposed method has been successfully applied to classify the lung nodules from the Lung Image Database Consortium (LIDC) database. The experimental results demonstrate a noteworthy efficacy of the proposition with sensitivity of 97.59% and 97.78% blind testing accuracy using only 12 discriminating features. Hence, the proposed methodology can be used to support the diagnosis pronounced by the radiologists interpreting manually the lung computed tomography images.