Computer-aided detection of tuberculosis among high risk groups: potential for automated triage

L. Hogeweg, A. Story, A. Hayward, R. Aldridge, I. Abubakar, P. Maduskar and B. van Ginneken

in: Annual Meeting of the Radiological Society of North America, 2011

Abstract

PURPOSE Tuberculosis (TB) screening programs are expensive because of, the large numbers of chest radiographs (CXR) that need to be read by human experts. The performance of a computer aided detection (CADx) system to detect TB on CXR is evaluated to determine its potential to triage images within a high throughput digital mobile TB screening program for high risk groups in London, UK. METHOD AND MATERIALS A large image database consisting of 47,510 CXR from 38,717 individuals was collected by the screening program between 2005 and 2010. In that period 120 screened patients were diagnosed with pulmonary TB. A set of 184 digital chest radiographs (DigitalDiagnost Trixel, Philips Healthcare, The Netherlands) from the screening program were selected to evaluate the CAD system on its ability to discriminate between TB proven and non-TB images. For training 89 images were used (69 consecutive non-TB images and all 20 culture proven TB images from 2006). The system was tested on the remaining 95 cases (67 consecutive non-TB images from 2006 and all 28 culture proven TB images from 2009). The research prototype CADx system (Diagnostic Image Analysis Group, Nijmegen, The Netherlands, Delft Diagnostic Imaging, Veenendaal, The Netherlands) was originally developed for analysis of CXR from high burden countries in sub Sahara Africa. It was retrained with abnormal regions that were outlined in the proven TB training cases. Diffuse abnormalities are detected by classifying small patches as normal or suggestive of TB inside automatically segmented unobscured lung fields. The probabilistic labels of the classified patches are then combined into one abnormality score for each image. The CADx system was evaluated on these scores using Receiver Operator Characteristic (ROC) analysis and specificity at 95% sensitivity. RESULTS The area under the ROC curve of the CADx system was 0.86. At a sensitivity of 95%, the specificity was 60%. CONCLUSION CADx can identify a large proportion of normal images in a TB screening setting at high sensitivity and has potential to be used for triage. CLINICAL RELEVANCE/APPLICATION Initial results demonstrate that the system could discard approximately 60% of images to potentially reduce the workload and costs of human readers while keeping a sensitivity of 95%.