Multi-class unsupervised classification with label correction of HRCT lung images

In this paper, we present an automated texture based unsupervised system for the classification of lung high resolution computed tomography findings in emphysema, ground-glass opacity, honeycombing and bronchiectasis. The classification techniques used in our study are based on cluster analysis of textural features. Variations of traditional K-means clustering are applied in the HRCT setting. A novel technique called label correction capable of segmenting "true" labelled pixel groups within the regions outlined by domain experts is presented. Label correction helps to "clean" the training data before supervised learning, and also provides more accurate evaluation on the testing data. The system was tested on 321 HRCT scans comprising varying diseases together with normal scans and successfully evaluated on the manually labelled scans by the doctors. In addition, the image segmentation results were visually validated by the radiologists.