Abstract
The study employs the gamma-selection enhancement to segment a radiograph of a hand and a wrist, and to build a bone age estimation (BAE) system. The proposed methods mainly include three steps: preprocessing, segmentation, and features analysis. First the preprocessing stag contains left hand cropping, background removal, orientation correction, and detection of phalangeal bone region of interest (PROI), and locating the position of epiphyseal/metaphyseal region of interest (EMROI) from local extremes. To choose the adaptive gamma parameter for equalization enhancement is one of the focal points in this thesis. Four quantitative shape measurements including misclassification error, relative foreground area error, modified Hausdorff distances, and edge mismatch via gamma-selection equalization are used to make the phalangeal assessment. By using comparison of error values, we can choose the adaptive parameters. A segmentation step is done for several methods consisting of adaptive two-means clustering algorithm, GVF snake, and local gray-level analysis segmentation. Comparing with other two methods, adaptive two-means algorithm with gamma-selection enhancement is requested for the later features analysis stage. We then extract three different sets of features, texture features, shape features, and geometric features, to build a series of membership functions for BAE. The effectiveness of the three sets of features is analyzed and a simplified procedure of feature selection is introduced to choose some representative features by using the correlation coefficients between the features and chronological age (CA). The results show that the BAE with the geometric features have better discrimination power than other two sets. The overall success rate of ROI extraction is about 90%, and the correct rate of assessing BA within 1.0 year errors and 1.5 years errors are 70.46% and 88.07% for female, and 73.34% and 86.97% for male.