Automatic detection of small lung nodules in 3D CT data using Gaussian mixture models, Tsallis entropy and SVM

Abstract

Lung cancer stands out among all other types of cancer for presenting one of the highest incidence rates and one of the highest rates of mortality. Unfortunately, this disease is often diagnosed late, affecting the treatment result. One of the hopes for changing this scenario lies in achieving a more precocious diagnosis of lung cancer through low-dose computed tomography, used as a screening method in risk groups of smokers or former smokers with elevated tobacco load. In order to help specialists in this search and identification of lung nodules in tomographic images, many research centers develop computer-aided detection systems (CAD systems) which are intended to automate procedures. This work has the purpose of developing a methodology for automatic detection of small lung nodules (with sizes between 2 and 10mm) through image processing and pattern recognition techniques. Some of these techniques are widely used in similar applications, as is the case of the region growing technique for segmentation of the pulmonary parenchyma. Other techniques, with more restricted application, are the Gaussian mixture models and the Hessian matrix for segmentation of structures inside the lung, Tsallis׳s and Shannon׳s entropy measurements as texture descriptors, and support vector machine to classify suspect regions as either nodules or non-nodules. The results achieved with the use of this set of techniques, applied to a sample with 28 exams from a public database, showed that small nodules were detected with a sensitivity of 90.6%, a specificity of 85% and an accuracy of 88.4%. The rate of false positives per exam was of 1.17.