Automatic detection of tuberculosis in chest radiographs using a combination of textural, focal, and shape abnormality analysis

Tuberculosis (TB) is a common disease with high mortality and morbidity rates worldwide. The chest radiograph (CXR) is frequently used in diagnostic algorithms for pulmonary TB. Automatic systems to detect TB on CXRs can improve the efficiency of such diagnostic algorithms. The diverse manifestation of TB on CXRs from different populations requires a system that can be adapted to deal with different types of abnormalities. A computer aided detection (CAD) system was developed which combines the results of supervised subsystems detecting textural, shape, and focal abnormalities into one TB score. The textural abnormality subsystem provided several subscores analyzing different types of textural abnormalities and different regions in the lung. The shape and focal abnormality subsystem each provided one subscore. A general framework was developed to combine an arbitrary number of subscores: subscores were normalized, collected in a feature vector and then combined using a supervised classifier into one combined TB score. Two databases, both consisting of 200 digital CXRs, were used for evaluation, acquired from (A) a Western high-risk group screening and (B) TB suspect screening in Africa. The subscores and combined TB score were compared to two references: an external, non-radiological, reference and a radiological reference determined by a human expert. The area under the Receiver Operator Characteristic (ROC) curve Az was used to measure performance. Additionally, the performance of an independent human observer was compared to the best individual subscore and to the combined TB score. For database A, the best performing subscores achieved Az = 0.827 and 0.821 for the external and radiological reference respectively, whereas in database B Az = 0.759 and 0.866 was achieved. Different subscores performed best in the two databases. The combined TB score performed better than the individual subscores, except for the external reference in database B, giving performances of 0.868 and 0.847 in database A and 0.741 and 0.899 in database B. The performances of the independent observer, 0.910 and 0.942 in database A and 0.755 and 0.939 in database B were slightly higher than the combined TB score. Compared to the external reference, differences in performance between the combined TB score and the independent observer were not significant in both databases. The combined TB score performed better than the individual subscores and approaches performance of human observers with respect to the external and radiological reference. Supervised combination to compute an overall TB score allows for a necessary adaptation of the CAD system to different settings or different operational requirements. I.