Statistical local texture analysis applied to computer-aided diagnosis in chest radiography

A system that automatically detects textural abnormalities in medical images will usually require three steps: 1) segmentation, that allows the selection of corresponding regions of interest; 2) texture feature extraction for each region of interest; 3) classification of regions, using the texture features and the ground truth from a training database. After classifying regions in the image, an overall abnormality score for the image can be computed. Such a system is described here, using active shape models for segmentation, a multi-scale texture feature extraction method based on Gaussian derivative filters and k-nearest neighbor classification. It is applied to two image databases of chest radiographs: one from a tuberculosis mass chest screening program and one from clinical chest films obtained in a general hospital.