Abstract
Emphysema is a chronic obstructive lung disorder that causes shortness of breath. Emphysema detection from computed tomography scans is normally done by analysing areas with low attenuation and texture variations in the lungs. To capture both intensity information and texture variations from the lung scans, in this study texture features are extracted using local binary patterns, Gabor filter bank, gray-level co-occurrence matrix and intensity information is extracted from gray-level histogram. The extracted texture and intensity features are inputted to metacognitive neural network to classify as normal, centrilobular emphysema and panlobular emphysema. During the training phase, samples are inputted one-by-one to the metacognitive neural network. The samples are added to the network if it has distinct knowledge otherwise samples are deleted. The study has been conducted on computed tomography scans available from Bruijne and Sorensen benchmark dataset. The performance of metacognitive neural network is compared with commonly used classifiers namely k-nearest neighbour, support vector machine, decision tree and also with similar works on emphysema classification. The comparison results clearly indicate the better performance of the proposed approach for emphysema classification with fewer data samples.
