Computerized analysis of interstitial disease in chest radiographs: improvement of geometric-pattern feature analysis

We have been developing automated computerized schemes to assist radiologists in interpreting chest radiographs for interstitial disease based on texture analysis and geometric-pattern feature analysis. In this study, we attempted to improve the performance of the geometric-pattern feature analysis, because the current classification performance with geometric-pattern feature analysis is considerably lower than that of texture analysis. In order to improve the performance in distinguishing between normal lungs and abnormal lungs with interstitial disease, we attempted to remove rib edges in regions of interest (ROIs) by using an edge detection technique, and also to reduce false positives by using feature analysis techniques. In addition, the effects of many parameters on classification performance were investigated to identify proper threshold levels, and subsequently the specificity of the geometric-pattern feature analysis was improved from 69.5% to 86.1% at a sensitivity of 95.0%. Using a combined rule-based method with texture analysis and geometric-pattern feature analysis plus the artificial neural network (ANN) method for classification, a high specificity of 96.1% was obtained at a sensitivity of 95.0%.