A screening system for the assessment of opacity profusion in chestradiographs of miners with pneumoconiosis

摘要

The aim of this study was to develop a screening system of chestradiographs of miners with pneumoconiosis. Chest radiographs were ofcoal mine or silica dust exposed miners participating in a healthscreening program. A total of 236 regions of interest (ROI) (166, 49,and 21 with profusions of category (shape and size) 0, 1(q), and 1(r),respectively) were identified from 74 digitized chest radiographs by twoB-readers. Two different texture feature sets were extracted: spatialgray level dependence matrices (SGLDM), and gray level differencesstatistics (GLDS). The nonparametric Wilcoxon rank sum test was carriedout to compare the different profusion categories versus that ofprofusion 0 (normal). Results showed that significant differences exist(at a=0.05) between 0 versus 1(q), and 0 versus 1(r) for 14, and 12texture features respectively. For the screening system, theself-organizing map (SOM), the backpropagation (BP), and the radialbasis function (RBF) neural network classifiers, as well as thestatistical k-nearest neighbour (KNN) classifier were used to classifytwo classes: profusion 0 and profusion 1(q and r). The highestpercentage of correct classifications for the evaluation set (116 and 20cases of profusion 0 and 1(q and r) respectively) was 75% for the BPclassifier for the SGLDM feature set. These results compare favorablywith inter- and intra-reader variability