Multi-classifier semi-supervised classification of tuberculosis patterns on chest CT scans

Classification of different textures present in chest CT scans of patients with pulmonary tuberculosis (TB) is of crucial importance for the success of ongoing vaccine and drug testing trials. In this paper, a new multi-classifier semi-supervised method (MCSS) is proposed that is trained with a small set of labeled examples and improves classification performance by sampling interesting samples from unlabeled scans based on uncertainty among a pool of classifiers. The interesting samples are added to the small labeled set with a label assigned by 'expert' classifiers. MCSS is applied to 20 scans of patients with proven TB for which a reference standard was obtained by a consensus reading. Another set of 35 scans was used without manual labels. The performance of MCSS is compared to conventional supervised classification and two other semi-supervised methods and shown to outperform all other methods.