Robust Computer-Aided Detection of Tuberculosis in Chest Radiographs Using Energy Normalization

PURPOSE The performance of computer-aided detection (CAD) algorithms for chest radiography can be influenced by variations in image data coming from different sources. Acquisition settings, detector technology and proprietary post-processing all influence the appearance of radiographs. We developed an algorithm to standardize the appearance of chest radiographs (CXRs) in order to remove these variations prior to image analysis and evaluated its utility for a CAD system aimed at tuberculosis (TB) detection. METHOD AND MATERIALS Three data sets of 200 digital CXRs were used: 100 normal / 100 abnormal cases from an Odelca DR system acquired in Zambia; 100 normal / 100 abnormal cases from a digital Atomed mobile X-ray system acquired in The Gambia; 127 normal / 73 abnormal cases from a Philips Digital Diagnost system acquired in the United Kingdom. Reference standard for suspicion of TB was set by an expert reader. To standardize the appearance of CXRs, the image is decomposed into frequency bands using hierarchical unsharp masking. In a training set the average energy (standard deviation) of each frequency band in the central part of the image is determined. Each energy band is scaled to this reference energy, and the input image is reconstructed from the scaled frequency bands. Subsequently the lung fields and mediastinum are segmented via pixel classification and the energy normalization is repeated for region containing the union of lung fields and mediastinum. Cases were processed by a CAD system (CAD4TB v3.07, Diagnostic Image Analysis Group, Nijmegen, The Netherlands) with and without applying the energy normalization method. This CAD system was trained with cases from an Odelca DR system. Performance was measured as area under the ROC curve (Az). Pairwise comparisons were made with bootstrap estimation, considering p<0.05 significant. RESULTS Without normalization, CAD4TB obtained an Az of 0.80, 0.61 and 0.47 for the data from Zambia, The Gambia, and the United Kingdom, respectively. With normalization, Az increased to 0.87, 0.80 and 0.84. Differences for the data from The Gambia and the United Kingdom were significant. CONCLUSION The robustness of CAD for detection of signs of TB on CXRs is improved by standardizing the radiographs prior to analysis. CLINICAL RELEVANCE/APPLICATION An automated reading system for CXRs that can be used reliably on data from any digital unit has great potential in TB screening and active case finding.